nschoe's labs Home Articles Haskell A.I. WebRTC About Contact RSS feed logo

Docker: Taming the Beast - Part V: Docker Networks


Hello again and welcome to Part V!
Part V… this begins to look like a real series of articles, I’m happy. And I’m happy that you’re still with me.
If you missed it, here is the Part IV, and in case you’re new here, you can start at the beginning here.

Small disclaimer

I’m sorry that this article took so much time, and I know you have been waiting for it: thank you for your emails.
As you will see, this article is very long (about twice as long as Part IV, which you surely recall was already quite big). This is because, as I will hint in this article, the topic is extremely important and this was crucial for me to get it right.
I hope I managed it and I hope I will definitely clarify everything about Docker Networks.
Have fun reading the article and thanks to all of you who supported me through email, this is highly appreciated!

Answering Part IV’s Question

Before we begin with Part V, if you remember correctly, I left a question without an answer in Part IV. It was at the end of section “Verifying the Read-Write Layer Conspiracy”.
We were checking container’s size. In a first example we had created an image with a certain size. Then we had added a 15MB file in it and then the container’s size grew by 15MB. When we deleted the file, the container’s size shrank by 15MB. Quite logic.

But then we had created a Dockerfile from the same image, but this time we created the 15MB file directly in the Dockerfile (as part of the image). But then when we deleted this file in a container, it did not shrink. Why was that?!

The answer, albeit a bit tricky is quite simple. In the first example, when we created the 15MB file in the container, the data from this file was put in the container’s RW layer (remember this is a layer that goes on top of the image’s layer, because an image is frozen, you can’t alter it). And so when we deleted the file, it was simply deleted from the container’s RW layer.

But in the second example, the file is being created in the image. So the 15MB file is part of the image. In other words, the data for this file is put in the image’s layers. So when we first instantiate a container, it’s RW later is empty. The total size is the same, though.
But since we cannot modify an image’s layers, we can’t simply delete the 15MB file like we did earlier (since it is in fact in the image’s layers!). But we still need the file to disappear, because we ran rm on it!

The way to solve this was to create a “shadow file” in the container’s RW layer. Basically, when you delete a file or a directory that is part of the image’s layers, a new, very small file is put in the container’s RW layer, that “hides” or “shadows” the target file. It’s like you cover the file and say “well, from now on, it doesn’t exist anymore”. And then when you run ls or the likes, the file is hidden / shadowed and it doesn’t appear to you.

Note that it is not simply hidden for your eyes when you run ls, it’s hidden filesystem-wise, it’s hidden for the actual system.
But since the actual bytes are not removed, it explains why the image’s size did not change.

That’s it!

Road To Awe

Last time, we saw how we could make part of our data persistent and bypass the unionFS mechanism. This was especially important if we did not want to lose data, and this allowed us to have truly ephemeral containers, i.e. containers that we can shutdown / destroy and recreate at any moment.
This is really important, and I have had lots of examples in the last couple of weeks. On the #docker IRC chan, there has been quite a number of people that came and asked variants of “is this possible to bind ports on a container without destroying and recreating it?”. This is not possible, the only solution is to effectively destroy the container and recreate it with the correct ports binding, and that should not be a problem!
Since containers should be ephemeral, all important data should be stored in Volumes: destroying containers should never be a problem.

Anyway, that was topic for Part IV, we’re in Part V now. So what’s this about?
It’s about the other most misunderstood Docker feature: Docker Networks. I don’t blame people because this feature is fairly new (when compared to the rest of Docker), but this has still been quite a while, so with the help of this article, I hope it will all be clear.

To make it extra short, the Docker Network is a feature that enables you to make your containers communicate together!

Excited?! Then read on!

The Basics

In this article, we will need to test network connectivity because we’ll be talking about networks. I have hesitated a long time between using standard Linux tools to test connectivity and making a small real example. To be honest I had even begun writing the article with the example, but I felt that it was too much information in a single article, and besides… Docker Networks is a very misunderstood feature and I would not want people to focus more on the example than on the real topic of the article.

So we will use netcat to test connectivity, which is the normal, standard Linux way. If you already know how to use netcat easily, this will allow you to focus entirely on the Docker networking aspect. If you do not, fear not, this is very easy to understand, I will explain everything, as usual and you’ll get to have one more tool in your toolbox!

And since I can already hear you complaining about the fact that we don’t make a real example, don’t worry: I am cooking an upcoming Part which will be a full article, entirely dedicated to making a slightly complex stack, real example.

Happy? :)

Let’s do this, I’m excited!

Why Do We Need Docker Networks Anyway?

If you’ve followed this series—or heard a tiny little bit about Docker—you have understood by now, that Docker is about isolation. It’s about running processes on the same Linux host, but each of these processes, “invisible” to its neighbors.
This works actually so well, that we run in a problem pretty soon when we’re trying to build something slightly more complicated than the examples we’ve seen up to here.

Let’s see this problem in action, right now.

As I said just above, we will use netcat to test connectivity. I will now make a very small introduction to netcat so that people who don’t know about it can follow. If you need a small refresher, read on, otherwise, you can skip it.

(Very) Small Introduction to netcat

netcat is a standard Linux tool that is used to test TCP and UDP connections. If you know telnet, it’s a sort of telnet on steroids, at least for the usage we will make.

Just to be clear, when I say “test connectivity”, I’m not talking about those speed-test you make to know how fast you can download and upload, no no, I’m talking about ports being opened and hosts being reachable.

So the easiest and quickiest way to do this is to have a (very) small server listening on the desired port on the desired host, and a (very) small client that tries to connect to the given host on the given port. That’s it. Really.
That’s all you need if you want to test connectivity. And netcat is perfect for this.

Just before we begin, know that the command nc is an alias for netcat. So if you read articles on Internet that has commands beginning with nc, it is netcat. Use whichever you want.

So how do we create a very simple server with netcat? It’s easy, you need to tell netcat that is should listen (and not connect), so this is done with the -l option.
Then you need to specify a port on which it should listen, this is done with the… -p option. See? Simple.

Mind that as a non-root user, you cannot create a server on ports 0-1024, so make sure you pick a port that is above.

For instance, let’s make netcat listen on port 1089:

$> netcat -l -p 1089

“What happened? My screen seems stuck.”. It’s a good sign, it means netcat succeeded in creating a server, and it is now listening on the port you specified.

Now, open another terminal, we’ll need to make a client. As a client, netcat needs to know two things: the host and the port to which it should connect to. So:

$> netcat localhost 1089

Again, nothing seems to happen: no feedback. But it there was an error, it would have displayed it. So now your client is connected. Try writing something in the client’s console, like "Hello", and press <Enter>.
You should see the string written on the server’s console, we now have confirmation that the two can communicate.

Alternatively, if you want feedback, use the -v option (for “verbose”):

$> netcat -v localhost 1089
localhost [] 1089 (ff-annunc) open

This way you have some feedback that the connection was successfully established.

Told you it was easy, right?!

This is what we will do all along this article, to test connectivity when we need it, so make sure you understand what’s going on.

So… What Was The Problem Again?

Yes, yes I’m coming to it!

So now that we have seen the basics of netcat, you will be able to understand when something works or doesn’t.
So, containers means isolation, as we’ve said before. But sometimes, it means too much isolation.

Let’s suppose for a moment that we are building a sort of webserver and we need a server (nginx) and a database (postgresql). As we’ve understood Docker’s basics, we will separate the two instances in two containers.

Let’s emulate this by creating two containers, we will name them accordingly, but we won’t really setup nginx or postgresql. Instead we’ll use netcat to create a server or connect to the corresponding port. This should be enough to highlight the problem.

Just before we begin, we have a small thing to do. Since we will be using netcat a lot in this article, we need an image with netcat installed; otherwise it will be a pain to re-install it everytime.
Up until now, we have only worked with the ubuntu and nginx images. So as not to change too many variables at the same time—which is often crucial to learn correctly—we will keep using ubuntu as a base image.

So let’s make a very small Dockerfile that creates a net-test image:

FROM       ubuntu:16.04
MAINTAINER nschoe<nschoe@protonmail.com>

RUN        apt update && apt install -y --no-install-recommends netcat inetutils-ping iproute2

Now build it with docker build -t net-test . and it’s done. Obviously starting from such a big image as ubuntu:16.04 (which is in the 600MB) just to have netcat is highly inefficient. But the goal here is to learn about networks.

Okay, back to our test.

Let’s begin with postgresql. By default, postgresql’s port is 5432, so let’s create a container and inside it, a netcat server listening on this port.

To emulate the container running postgres, first create a container from net-test:

$> docker run --rm -it --name fake-postgres net-test bash
root@6995c50eaff3:/# netcat -l 5432

Okay, so our fake postgres container is running. Let’s now create a fake nginx container, which will connect to the fake postgres:

$> docker run --rm -it --name fake-nginx net-test bash

Now how to connect this fake-nginx to the fake-postgres? We saw earlier that the syntax was netcat <host> <port>. We know the port, it’s 5432. But what about the host?

First idea: localhost, after all, why not? Containers are on the same host, let’s try:

root@9e09afc2d19a:/# netcat localhost 5432
localhost [] 5432 (?) : Connection refused

Okay. That doesn’t work.
And after all this is to be expected: localhost inside different containers resolve each to their container, otherwise we could not really say they were isolated.
Careful readers might have a suggestion: you may have seen the PS1: root@9e09afc2d19a.
Usually, the part after the @ is the hostname. And sure enough, if you try to run hostname in the terminal, you get just that:

root@9e09afc2d19a:/# hostname

So we might try this.
But the two containers have a different hostname, so which one to try? Let’s try both, to be sure:

root@9e09afc2d19a:/# netcat 9e09afc2d19a 5432
9e09afc2d19a [] 5432 (?) : Connection refused

root@9e09afc2d19a:/# netcat 6995c50eaff3 54321
6995c50eaff3: forward host lookup failed: Host name lookup failure : Resource temporarily unavailable

So the first had no chance of succeeding: in the fake-nginx container, I tried to connect to host fake-nginx, which is localhost, and we’ve just seen that it failed. The second attempt however, made much more sense; we tried to connect to the host of the fake-postgresql container, which should work.

It takes time, but then eventually it fails. And this precisely, is our problem: containers are isolated, even on the network stack.

So this is what starts this whole article: how do I make these two (or more) containers communicate?

Now I hope I convinced you that we do have a problem, and now let’s see if I can help you understand how to fix it!

Two Separate Use-Cases

So we have experienced the containers’ network isolation, and we have understood that this might be a problem when we actually need communication. The rest of this article will be about how to break that isolation in a controlled way.

As often with Docker, there are more than one way to do it, and as it is always the case, they correspond to different use-cases.

Here, when it comes to network isolation, we will talk about two use-cases. This might be a little early, but I’d like you to take a minute (you know, we have talked about this… do try to think about this for a minute, see if you can find ideas) to find the two use-cases.
Here is a piece of advice: it deals with whom containers have to talk to.

I hope you tried, let’s see if you got that right!

The first use-case is simple: this is the one we’ve just demonstrated; it is when two (or more) containers need to talk to each other. It might be that you have a webserver (like nginx) that needs to talk with a database (postgresql), maybe we can even imagine a caching stage (like redis). All these containers need to talk to each other, and so it is a first use-case.

The second use-case if when the host needs to talk with a container. You might have forgotten about this, but if Docker can be used to ship applications in a controlled environment for a client, or be used to deploy a great number of application stacks, it can also be used to containerize one instance of a program for use on your host.

Since your host (i.e. the computer running docker) is not a container, there are different mechanisms to make it communicate with the containers. We will see both cases.

First Use-Case: Communication With the Host; Binding Ports

We will start with how we can make the host communicate with containers, as it is generally the first use-case people have.

Suppose you are trying to containerize a web server, like apache or nginx for your host (for instance, to serve your local, self-hosted website(s)). This is a classic situation. You might be interested in trying a new version of your webserver but are not ready to upgrade the version on your host, or you might be trying to switch from one to another, or you simply want to keep it isolated from the rest of your environment.

Anyway, this is what we are trying to do. The same kind of problem that we saw in the very first paragraph will arise: we need a way to access our container, as it is it, which is running the webserver and has the data in it.

The situation is really simple: when we create a container for apache or nginx, and we’re done configuring it, this server inside will listen to ports 80 and 443. But Docker is good at isolation, and thus he created a nice, little isolated network environment for the container. Meaning that the container has its own new (virtual) network interface, and through it it communicates (or not) with the rest of the world.
Remember: the goal of Docker toward your containers, is to make them believe they are running alone in the system, and control everything it can about it, including its Network I/O.

As a reminder, and if you want to see it with your own eyes, first stop & destroy every container in your host. When you have no other (check with docker ps -a), let’s look at the network interfaces your host has:

$> ip -oneline link show

1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state UNKNOWN mode DEFAULT group default \    link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
2: wlp4s0: <NO-CARRIER,BROADCAST,MULTICAST,UP> mtu 1500 qdisc mq state DOWN mode DORMANT group default qlen 1000\    link/ether 34:02:86:2b:79:9c brd ff:ff:ff:ff:ff:ff
3: docker0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP mode DEFAULT group default \    link/ether 02:42:70:6d:78:ff brd ff:ff:ff:ff:ff:ff

Here we can see 3 networks interfaces: the local loop lo, my WiFi connection wlp4s0 and the docker0 virtual network interface that docker created when your installed it.
If your computer has other network interfaces, you’ll see them here.

Now let’s run a container with docker run --name test -itd nginx bash, and run this command again:

$> ip -oneline link show

1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state UNKNOWN mode DEFAULT group default \    link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
2: wlp4s0: <NO-CARRIER,BROADCAST,MULTICAST,UP> mtu 1500 qdisc mq state DOWN mode DORMANT group default qlen 1000\    link/ether 34:02:86:2b:79:9c brd ff:ff:ff:ff:ff:ff
3: docker0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP mode DEFAULT group default \    link/ether 02:42:70:6d:78:ff brd ff:ff:ff:ff:ff:ff
11: veth9170aac@if10: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue master docker0 state UP mode DEFAULT group default \    link/ether 42:a7:d9:db:ca:17 brd ff:ff:ff:ff:ff:ff link-netnsid 0

Here you can see a new (virtual) network interface named veth9170aac@if10. the "veth" prefix stands for "virtual ethernet", and it represents the (ethernet) network interface of the container we have just created.

The goal of all of this, is that for the container, it has its own network interface. This means that you are free to use any port inside the container, as they are separated from the host’s network interface. By default then, “all ports are free”.

The TL;DR of this is that port N for the container is not really port N on your (host) machine. Actually it is not even a port at all for the host machine (this actually involves iptables rules, more on this later in this article).

Okay, So How do I Make My Host and My Container Talk to Each Other?

You bind ports. That’s how.
What that means, is that you bind (as in “connect”) a port on one of the host’s network interface, to a port on the container’s network interface.

Concretely that means, that everything going through one of these ports is going out on the other. You can link the host’s port 444 to the container’s port 443. This way, when you’re accessing the host’s 444 port, your requests are actually forwarded to the container’s 443 port, and vice-versa.
If the host’s interface on which the port is bound has external (Internet) connectivity, this is how you can grant people outside access to your container’s services.

As often with Docker, there are several syntaxes, let’s review them here.

Explicit Mapping

When binding ports with explicit mapping, you define exactly which ports on the container are bound to which ports on the container. This is obviously done at run time (i.e. docker run). For that, you use the -p (this is lowercase p). The syntax is: docker run -p <hostPort>:<containerPort>.
For instance, to bind the container’s 443 port to the host’s 444 port, you would do docker run -p 444:443 <image>.

Also some people find it counter-intuitive (I don’t), make sure to remember that the first is the host’s port and the second is the container’s.
Of course, nothing prevents you from using the same port, so you can run docker run -p 443:443. This is used when you want to containerize a service on the host.

This goes without saying, but once you bind a port on the host, you cannot re-use that port. This is not docker, this is classic networking, right? You cannot have two applications listening on the same port, this is why I said that explicit port mapping is used when you want to containerize a service on the host: usually you have one nginx instance running on your computer and it has responsibility of the port 443.
Well, here this is the same: you have one container running nginx and you’ve bound the ports 443 of both. If you want another nginx instance, your need to bind this second instance’s 443 port to another port on the host.

Random Mapping

This is the opposite of explicit port mapping. In this scenario, you want to bind the container’s port to a port on the host (understand: you don’t know / care which one). This is used when you want to have several instances of the same application, and you want to bind the ports. But you don’t want to / cannot specify the ports every time.

The syntax is: -p <containerPort>. For instance: -p 443. This is basically saying “I want the container’s 443 port bound to the host, use a random port on the host”.

This is cool because you don’t have to remember and specify the ports on the host on which you want to bind (so this is easier to automate), but the drawback is obviously that you need a way for the application that needs to contact your container to discover the port it needs to connect to.

A use-case for this can be for debugging. For instance, if you have a setup that includes spawning database containers, you can bind the database server’s access port randomly on the host. Most of the time, you would not need it, so it doesn’t really matter, but when comes the days when you need to debug, find the port (we’ll see how at the end of this section) and connect to it.

Note that when binding to random ports, the random ports are chosen in the 32,7xx range.

Specific Interface

By default, when you bind ports (whether specifically or randomly), it is bound on interface, which means “all interfaces”. What this means is that if you bound your nginx’s 443 ports to the host’s 443 ports (for instance), then this 443 port will be routed to your container’s, whatever the interface used: eth0, eth1, wlan0, localhost, etc.

It is common for servers to have several network interfaces (meaning several Ethernet connections for instance, several network cards) with different IPs. If you don’t specify otherwise, Docker will bind ports on all of these interfaces.

This can be useful (for instance in the case of binding commonly-used ports, like 80, 443, etc.). But it might also be very dangerous: if we take the previous example, where I said that if you had several database containers, you could bind the server port (5432 for PostgreSQL for instance) randomly so that you could debug when needed, then it is dangerous. Doing this would essentially open a port on the outside that is routed to your container’s database.
Normally you have set up well-protected users and passwords to access the database, but still, this is dangerous.

Let’s see a quick example of this.

First, find one public IP of your host with ip addr show, under the inet section (for ipv4 address) or inet6 (for ipv6 address).
This is your “public ip” (I’m using quotes here, because if you’re doing this on your own computer, at home or at work, you’re most likely behind a router NAT-ing your connection, so you’ll have a local IP on the network ; only if you’re doing this on a remote server will you have a real “public” IP).
But it doesn’t matter, it works exactly the same, just consider it to be your IP to be public, the one unknown users on the Internet use to connect to your host).

$> ip -o addr show
1: lo    inet scope host lo\       valid_lft forever preferred_lft forever
1: lo    inet6 ::1/128 scope host \       valid_lft forever preferred_lft forever
2: wlp4s0    inet brd scope global dynamic wlp4s0\       valid_lft 3332sec preferred_lft 3332sec
2: wlp4s0    inet6 fe80::3602:86ff:fe2b:799c/64 scope link \       valid_lft forever preferred_lft forever
3: docker0    inet scope global docker0\       valid_lft forever preferred_lft forever
3: docker0    inet6 fe80::42:1fff:fe82:d2ea/64 scope link \       valid_lft forever preferred_lft forever

As you see on the 3rd line (first entry prefixed 2:, I’m using my wlp4s0 interface here, and my “public ip” is:

Then create a container from image postgres, bind the port 5432 (in PostgreSQL, this port is the port used to connect to the database server, so this is where you can take control of the database if you manage to connect).
You are most likely to already have port 5432 listened to on your host, so if you can afford stopping the postgres on your host for a couple of minutes, do it: sudo systemctl stop postgres, otherwise, in the next command, change the port on the host on which you bind the container’s 5432 port.

$> docker run --name port-binding-test --rm -it -p 5432:5432 postgres

Let it start, and then, it’s time to inspect the ports.

First, on localhost:

$> nmap localhost
Starting Nmap 7.40 ( https://nmap.org ) at 2017-03-08 09:09 CET
Nmap scan report for localhost (
Host is up (0.000058s latency).
Not shown: 999 closed ports
5432/tcp open  postgresql

Nmap done: 1 IP address (1 host up) scanned in 0.04 seconds

So nmap (which is a program used to scan opened ports on an IP) reports that the port 5432 is opened (here you’ll have your other ports, if you chose to bind randomly or if you manually specified another port). Let’s check if we can connect to it:

$> netcat -v localhost 5432                                                                 ~
localhost [] 5432 (postgresql) open

And yes, this works.

Now let’s run this nmap again, but on the host’s public interface (this is what a user would get if he ran nmap on your host’s IP address):

$> nmap

Starting Nmap 7.40 ( https://nmap.org ) at 2017-03-20 08:35 CET
Nmap scan report for
Host is up (0.000071s latency).
Not shown: 999 closed ports
5432/tcp open  postgresql

Nmap done: 1 IP address (1 host up) scanned in 0.12 seconds

Depending on the services running on your host, you might have other ports opened, so more lines like this. But the important part for our example is the line 5432/tcp open postgresql.

So we have confirmed that, when using the -p option to bind a port with the format: -p <host-port>:<container-port>, this port was bound for every interface.
And this is not always what we want (think again about opening a connection to our postgresql database to debug ; or suppose your host has a network interface on which there is no external connectivity but is used for LAN connections with other hosts, etc.).

The docker developers have again thought about this and provided us with a way to specify the interface when we want to bind ports, the syntax is: -p <ip-of-interface>:<host-port>:<container-port>.

We’ll try it again with our postgres example, first, shut the container from the previous example down (and remove it if you did not use the --rm flag).

This time, we will try to bind the 5432 port only to the localhost interface, which corresponds to the use-case we described earlier: leave a port opened so that if we need, we can have direct access to the database, but we want this to be possible only from the host itself (for someone logged in via SSH) and not for the whole world to see!

$> docker run --name port-binding-test --rm -it -p postgres

Notice how we prefixed our ports numbers with which is the localhost’s IP address. Let it start for a couple of seconds, and now, let’s run our nmap commands once again.

First, on localhost, we expect the port to be opened:

$> nmap localhost
Starting Nmap 7.40 ( https://nmap.org ) at 2017-03-20 08:43 CET
Nmap scan report for localhost (
Host is up (0.000066s latency).
Not shown: 999 closed ports
5432/tcp open  postgresql

Nmap done: 1 IP address (1 host up) scanned in 0.05 seconds

The output has not changed: the port is opened (this is obviously what we expected, since we bound the port to the localhost interface).

Now, let’s run nmap on the public IP, and this time we expect the port not to be opened:

$> nmap
Starting Nmap 7.40 ( https://nmap.org ) at 2017-03-20 08:44 CET
Nmap scan report for
Host is up (0.000069s latency).
All 1000 scanned ports on are closed

Nmap done: 1 IP address (1 host up) scanned in 0.32 seconds

Ah ah! Now the port 5432 is closed! Which is exactly what we wanted. (Note: as before, depending on the services installed and running on your host, there might very well be other ports opened. Please understand that here we don’t seek "All 1000 scanned ports on are closed", we simply seek the line 5432/tcp open postgresql not to be here!)

So this is a very handy tool, because now you can bind ports on the desired interfaces and have a very fine granularity of the way you expose your services.
Note that if your host has more than 2 interfaces (let’s say 5) and you want to bind the port on 3 of them but not the other two, you need to use the -p option three times: -p <ip-1>:5432:5432 -p <ip-2>:5432:5432 -p <ip-3>:5555:5432. Here I have bound the postgresql port 5432 to the same port number 5432 of the first two interfaces and on the port 5555 of the third interface.

Note that if you want to have specific interface and random port mapping, then the syntaxt is as follow: <interface-ip>::<port>. Note the :: in the middle (two :). This is basically the same as above, but with the “host” part omitted.
It looks weird, but that’s the way to do it.

Imagine the possibilities: you can bind some ports of your containers to the tun0 interface of your host and this gives access to the ports only through the VPN tunnel, or you can bind ports to the internal Ethernet interface to allow fast communication intra-datacenter.


Range of Ports

This is all very nice and all, but there remains one tricky part. Suppose you are using an application that spawns ports randomly in a specified range and you want to bind them to the host, how do you do?
With what we’ve seen so far, it is possible but very cumbersome: we would have to bind all ports manually. What if there is a hundred?

A common example of this situation is if you are dockerizing a WebRTC gateway for instance: it chooses a random UDP port for each client.
More generally this will happen every time you have a server that spawns a worker on a random port for each new connection.

What we are looking for now is for a way to bind a range of ports from the container to the host.

And lucky for us, there is a way, the syntax is as such: -p hm-hM:cm-cM where:

  • the - sign denotes a range
  • hm is the lower port of the range for the host
  • hM is the upper port of the range for the host
  • cm is the lower port of the range, for the container
  • cM is the upper port of the range, for the container

For instance, if you want to bind all ports between 2000 and 2100 in the container to ports between 5000 and 5100 on the host, you will write:

$> docker run --name port-binding-test --rm -it -p 5000-5100:2000-2100 postgres

If you want to check the ports bound for a given container, you have two possibilities: either docker ps and there’s a column PORTS, or you can use the dedicated command docker port <container-name>

In our case:

$> docker port port-binding-test

2000/tcp ->
2016/tcp ->
2047/tcp ->
2062/tcp ->
2063/tcp ->
2079/tcp ->
2085/tcp ->
2007/tcp ->
2033/tcp ->
2072/tcp ->
2073/tcp ->

The only caveat though is that the two ranges must be equal: you can’t have something like -p 5000-6000:2000-2100 which would means “bind ports 2000 to 2100 to whatever ports between 5000 and 6000”. I hope this will be fixed, but for now, you must make sure your ranges are of equal length.

Conclusion on Binding Ports

That’s it for use-case #1!
You know everything you need to know to establish communication between your host and your containers, network-wise. Just to sum up what we have just seen: we have understood that docker, being so good at isolation actually isolated the network stacks of our containers.

Meaning as far as the containers are concerned, they have their own network interface (usually eth0) which is bridged to one of the host’s network interface (more on this later) to provide external connectivity (“access to Internet”).

But docker hides this from the host, which means that, by default, they have no way to communicate: the host simply cannot talk to the container and vice-versa (the container doesn’t even “know” that there is a host!).

But sometimes you want to contact your containerized processes from the host: so you need to bind a port. This is literally just a way of “piping” network packets on one of the host’s port to one of the container’s.

This binding of port is useful for examples such as:

  • when you want to dockerize a host’s service, for instance if you want your apache or nginx webserver dockerized (meaning you don’t want an apache or nginx running “vanilla” on your host anymore: you want to isolate it).
    In such a case, you need to bind the ports to their host’s equivalent (so 80:80, 443:443, etc.). This is because when you want to containerize the service, you expect the service to work “as before”: applications used to contact your host computer on its 80/443 port, it should not change anything for them. So you simply, brainlessly pass the traffic incoming on your host’s 80 or 443 ports to your container’s 80 and 443 ports.
    Obviously this is true for any ports of any containerized service: you can containerize your database instance (postgresql, MySQL, etc).
    You can customize your redis cache instance, etc.
    Note that you can containerize any service you want: it doesn’t have to do with the network! But if it does happen to have something to do with networks, then make sure to bind ports.
  • when you want to be able to contact a dockerized application, for debugging, logging, maintenance, etc. This usually happens with applications for which you have several instances. In this case, you will usually use random port mapping (or if you have only a few of them, you can always specify the ports yourself, but past 4-5 instances it becomes cumbersome).
    Usually with such a setup, you tend to use random port mapping but on a specified (often private, non-externally reachable) network interface (maybe a local network interface on your host, or a VPN tunnel?)

We have just concluded a big part and it was full of important information. Just so you know, I estimate that a solid 30% of the questions I answer on #docker are about Docker Networks, and a lot of them pertain to this topic of making a container communicate with the host. You absolutely need to understand this.
I hope I made this crystal-clear, and please, if anything is still unclear, . I will answer, really. It’s paramount that you understand this.

I suggest you take a break now, because you need to digest everything you’ve just read. And as usual, I mean it: in the next section, we will begin use-case #2, and when it comes to Docker Networks, people always, and I mean always get it wrong. They always mix the two use-cases, so don’t read the two cases one right after the other, because in your brain, you will join them and they will become this big mess of “Docker Network-ish”. It’s important that you understand there are two topics in Docker Networks.

Second Use-Case: Make Containers Communicate Together

Hi back!

I hope you did take a break, and that you are ready for new concepts now, because if you thought this one was going to be easy, well… you’d be right, because I am going to make my best to explain in a easy-to-understand manner.
But seriously, this is going to be a difficult and complex topic (but so interesting!), so pay attention and don’t skip steps.

I may have already said it earlier, but Docker Networks is, I think, the #1 feature of Docker. This is the most fantastic thing Docker brought to us!

From the first section, we have seen how to break the isolation between a host and a container. This is very interesting and important: we have seen that we can open ports between containers and (some of) the host’s network interfaces.

The second use-case, which is equally (if not more!) important (and cool!) is to make two containers (or more!) communicate together, without the host’s intervention.
I insist on this last part: too often, way too often I see people use the first method we saw (opening ports between containers and the host) to try and make two containers communicate together. Bah! After you read this I don’t want you to even think about this.
This is wrong at so many levels.

No here, we will see a method, no… the method to make containers communicate together.

If you’re ready, let’s proceed!

The New, Powerful And Appropriate Way: Docker Networks!

From the last paragraph that you read, you should have an intuition of what Docker Networks are, and this is cool, because that’s exactly what I intended. If you don’t, don’t worry: I will explain.

Simply put, a Docker Network is comparable to a LAN or custom WiFi network. You create a Docker Network and connect containers to the network. The magic is that containers inside a same network can see each other, and thus talk to each other.

The most important aspect (the one I really want to insist on), is that Docker Networks are separate things from containers. On one hand you have your containers, on the second hand you have your Docker Networks. You choose to connect containers to Docker Networks or not, and it’s not definitive: once your container has joined / connected to a Docker Network, you can just as easily disconnect it.
In the beginning, I found that it helps a lot to think about Docker Networks as WiFi networks (for the analogy, of course). It helps understand, for instance, that contrary to --link, there is no “precedence”: container #1 is connected (or not) to the Docker Network, and this is also true for container #2.
But the order in which you connect them has absolutely no meaning: the same way that you don’t case if your smartphone is connected to your home’s WiFi network before or after your laptop. Really, this is it.

Another important difference with --link is that since containers and Docker Networks are two completely separate entities, you don’t have those leaked environment variables anymore: again, think of WiFi networks. Just because you connected your laptop to your WiFi network doesn’t mean every other laptop on the network can access all your data, and much less your environment variables.

And just to make sure, I’ll repeat it: connecting containers to a Docker Network breaks the isolation on the networking stack. It just means that your containers will be able to see and talk to each other. But let’s make it clear that no data is shared, no isolation is broken on the filesystem stack etc. You still have totally independent containers.
And this is the magic of Docker Networks: they break the isolation just enough, exactly like we need it to.

Let’s talk about Docker Networks more in depth!

Basic Commands

Commands that deal with Docker Networks are docker network commands. It obviously accepts subcommands that we will see now.
The very first thing we can do with Docker Networks is to create one! This is easy enough, it’s done with docker network create. This command takes some options and the name of the network. So let’s do it:

$> docker network create MyFirstNetwork

Like everything in Docker, a Docker Network has a unique identifier, which is this weird returned hash. We don’t really need it, as we will mostly deal with names. But, as with every Docker command, know that in place of names, you can use this ID if ever you need to (by the way, your hash will obviously be different than mine!).

So the fact that docker returned a hash means the creation was okay. How can we check this?

By listing the networks, of course. This is done with docker network ls, so let’s do it now and see our network:

$> docker network ls
NETWORK ID          NAME                DRIVER              SCOPE
b2c78b7b2ec6        MyFirstNetwork      bridge              local
40670738e633        bridge              bridge              local
2683c3824f8f        host                host                local
a84e731c4387        none                null                local

Wait, what?!
The output above is what you should see if you haven’t played with Docker before (or at least with Docker Networks).

So even though we have only created one Docker Network, we can see that we have four already, what’s going on? I’ll be explaining in a couple of paragraphs, bear with me for now. Right now, we’re interested in seeing that MyFirstNetwork has been created.

Now that we have a Docker Network on hands, we might want to try and look a bit at it, because so far we have only just seen its name, which is already something, but quite limited. There is the docker network inspect command, that can, as its name implies, inspect and give some more thorough information about a network.

$> docker network inspect MyFirstNetwork
        "Name": "MyFirstNetwork",
        "Id": "b2c78b7b2ec68ef1fe702be331aa9cd3f3e1733039b8a064c702652439e1b20b",
        "Created": "2017-04-24T08:51:55.089418851+02:00",
        "Scope": "local",
        "Driver": "bridge",
        "EnableIPv6": false,
        "IPAM": {
            "Driver": "default",
            "Options": {},
            "Config": [
                    "Subnet": "",
                    "Gateway": ""
        "Internal": false,
        "Attachable": false,
        "Containers": {},
        "Options": {},
        "Labels": {}

There is quite a lot of information here, and we will come back to it a bit later in this article, when I have introduced the important notions.
You can already recognize the Name of the network as you defined it on the docker network create command, the Id, which is the hash docker returned when you created the network. If ever that matters to you, you also have the date at which the network was created.

On simple thing we can talk about right now, is that, as you can see, by default, IPv6 is not enabled in docker networks. This means that —once we learned how to do it— when we have containers using this network, they can only talk with IPv4. To enable IPv6 in a network, you need to add the --ipv6 option when running docker network create. Like this: docker network create --ipv6 <network-name>.

The next logical thing we can do is remove the network, this is done with docker network rm. So let’s try it:

$> docker network rm MyFirstNetwork

When a removal is successful, docker prints back the name of the network you’ve just removed. You can confirm that the network doesn’t exist anymore with docker network ls.

Okay, so we’ve got the basic commands, now it’s time to let the fun begin!

The Principle Behind Docker Networks

Up until now, when we have created containers, we were in a situation similar to this:

Fig 1: containers exist separately, no communication between them
Fig 1: containers exist separately, no communication between them

Now, we want to sort of “group” containers together, to allow them to communicate and see each other. We are going to use Docker Networks to do this. For the vocabulary part, know that we say connect a container to/in a Docker Network (and obviously the converse is disconnect).
I know you are probably eager to learn more about Docker Networks and start playing right away, but there’s a reason I’m going slow: as I said in the very first Part of this series, this is typically an area where people tend to read fast (because that’s new an exciting), think they understand, skip on sentences and they find themselves with a vague notion of understanding but without the gotchas and nuances.

So here, make sure you don’t mix connect which is when a container connects to a Docker Network (so between a container and a Docker Network) and link which is the old-fashioned way, and which happened between two containers.

When we connect a container to a Docker Network it becomes part of this Docker Network’s network stack and the container gains a new (virtual) network interface. As far as the container is concerned, everything happened as if you had installed a new network card and connected this new network card’s Ethernet connector.

One container connected to a Docker Network in itself is not very interesting: there’s not much added behavior. The fun begins when you connect a second container to the Docker Network. And that is also where you begin to see the advantage of Docker Networks over the old-fashioned --link: it’s much more flexible: you have a Docker Network on a side, you have several containers on the other side, and you can dynamically connect and disconnect some (or all) of the containers to the Docker Network. And all of this, without affecting the other containers!

And what is even more powerful is that you can connect one container to several Docker Networks! The container will simply see a new (virtual) network interface. Exactly as the processes running on your computer would see a new network interface if you (physically) connected a new network card to your computer.

Practice Time!

That’s a lot to digest, I know, so we’re going to illustrate all of this.

Let’s start from scratch: no containers, no Docker Networks, nothing. Blank canvas.

Now we create one Docker Network, called CustomNetwork:

$> docker network create CustomNetwork

Let’s make sure it is created:

$> docker network ls

NETWORK ID          NAME                DRIVER              SCOPE
9aedc6a27184        CustomNetwork       bridge              local
9dbd76505eed        bridge              bridge              local
2683c3824f8f        host                host                local
a84e731c4387        none                null                local

Don’t worry if you have others (in particular the MyFirstNetwork we created earlier), it won’t be a problem. All that matters is that we have our CustomNetwork network.

Now, we are going to create three containers that we will simply call one, two and three to keep things simple.

$> docker run --rm -itd --name one net-test bash

$> docker run --rm -itd --name two net-test bash

$> docker run --rm -itd --name three net-test bash

(Make sure to keep the 3 terminals open, because we used option --rm to remove the containers once they exit).

Let’s check (for sanity’s sake) that we do have our containers:

$> docker ps --format "{{.Names}}"

Okay, perfect.

So basically, we are in this situation:

Fig 2: Our CustomNetwork Docker Network and the three containers one, two and three
Fig 2: Our “CustomNetwork” Docker Network and the three containers “one”, “two” and “three”

Remember what I told you about the story of container gaining a new (virtual) network interface? Well we’re going to check this.

Let’s enter container one and check its network interfaces, as a reminder, you enter the container with docker exec -it <container-name> bash (or replace with sh is bash if not present).

Once inside, we will list network interfaces with ip link show (or the caveman can use ifconfig if they want ^^):

$> ip link show
1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state UNKNOWN mode DEFAULT group default
    link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
6: eth0@if7: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP mode DEFAULT group default
    link/ether 02:42:ac:11:00:02 brd ff:ff:ff:ff:ff:ff

I will let you verify that the same command run in containers two and three will yield similar results. So what we see here is that we have two network interfaces: lo which corresponds to the local loop (localhost) and eth0, which… you should consider as the “default network interface” for now.
I will explain the truth about it a bit later in the article, for the moment, just admit that the container has a default network interface.

Okay, so you can exit this terminal (again: make sure to keep the first three terminals opened, not to terminate the containers).

Right now, we have one Docker Network (CustomNetwork) and three containers, totally unrelated. It’s time to unleash the power of Docker Networks!
We are now going to connect container one to the Custom Network.

This is done with the docker network connect subcommand. This subcommand takes the name of the network and the name of the container to connect.
A small warning on this command, some people find the order of the arguments weird (I don’t), but it’s the network name first, and then the container’s name.

$> docker network connect CustomNetwork one


Okay there was neither smoke nor flashy lights, but we have just done something interesting. Granted, docker doesn’t provide you with a lot of output, but docker network commands are part of this family of “shy” commands: when they don’t tell you anything, it means it went all right. They start talking where some bad things happened.

So we still need some confirmation that what we’ve just done was actually useful.

We can start by inspecting the network and confirm that the container is now connected to it:

$> docker network inspect CustomNetwork
        "Name": "CustomNetwork",
        "Id": "9aedc6a27184d02d488c79f85eb185e6e2bf841627b03449f42d110c2f7eb4f3",
        "Created": "2017-04-24T19:00:25.165477676+02:00",
        "Scope": "local",
        "Driver": "bridge",
        "EnableIPv6": false,
        "IPAM": {
            "Driver": "default",
            "Options": {},
            "Config": [
                    "Subnet": "",
                    "Gateway": ""
        "Internal": false,
        "Attachable": false,
        "Containers": {
            "0be487e8b021b720dd657496c1ae92450c28de8ef9f1b16f06c72fe057ea567e": {
                "Name": "one",
                "EndpointID": "c298e545a4b68da5771be1504b3d0a5f2435afda425b3be3d47d0ff7d8f6582a",
                "MacAddress": "02:42:ac:13:00:02",
                "IPv4Address": "",
                "IPv6Address": ""
        "Options": {},
        "Labels": {}

The interesting part is the Containers:

"Containers": {
            "0be487e8b021b720dd657496c1ae92450c28de8ef9f1b16f06c72fe057ea567e": {
                "Name": "one",
                "EndpointID": "c298e545a4b68da5771be1504b3d0a5f2435afda425b3be3d47d0ff7d8f6582a",
                "MacAddress": "02:42:ac:13:00:02",
                "IPv4Address": "",
                "IPv6Address": ""

This section lists the container that are currently connected to the Docker Network we introspected, and we see that a container with "Name": "one" is connected, so that’s a first indication that it worked!

We can also re-enter the container, and check its network interfaces: it should have gained a new one, as we mentionned before. Let’s check this now:

$> docker exec one ip link show
1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state UNKNOWN mode DEFAULT group default
    link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
6: eth0@if7: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP mode DEFAULT group default
    link/ether 02:42:ac:11:00:02 brd ff:ff:ff:ff:ff:ff
12: eth1@if13: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP mode DEFAULT group default
    link/ether 02:42:ac:13:00:02 brd ff:ff:ff:ff:ff:ff

Two comments here:

  1. usually, we use docker exec -it one bash and then we get a shell prompt and type our command(s) (here ip link show).
    What you must understand is that docker exec is a command to execute a command inside a container: so you supply the container’s name and the command to execute. What we usually do is execute the command bash, and since it needs a tty, we use option -t and since we need interaction, we use option -i (which, concatenated gives -it).
    But docker exec is really meant to execute whatever command you want, so this time, I took a shortcut and directly ran the command I was interested in, i.e. ip link show. Since this time it doesn’t need a tty and it doesn’t need interaction, I ommited the options -i and -o.
    So really, this should not confuse you; from now on, I’m going to use either interchangeably!
  2. back to our Docker Network topic: we can see here that we now have three network interfaces: lo and eth0 which we already had, and the new eth1. As expected, the container now sees a new network interface, because as far as it is concerned, we have just added a new network interface.

As a last confirmation (actually I’m doing all of these steps to give you closure: by showing you all the components on what is involved, I am trying to show you that everything is logic and there is absolutely no “black, magic box”), we are going to get the container’s IP address.

What we have just done is connect our one container inside a Docker Network: CustomNetwork. Since it joined a network, it must have an IP address inside this network. I’m going to show you two ways to get it, which should confirm what is going one.

The first way is to get the container’s IP address by inspecting the Docker Network. If you look closely at the Containers section from earlier:

"Containers": {
            "2deaab87b3f99c40c245c8e77480656d93cd14c55fa34229e6c062fa74d8e035": {
                "Name": "one",
                "EndpointID": "c298e545a4b68da5771be1504b3d0a5f2435afda425b3be3d47d0ff7d8f6582a",
                "MacAddress": "02:42:ac:13:00:02",
                "IPv4Address": "",
                "IPv6Address": ""

We can see that for the container one, there is a IPv4Address, whose value is (the value is likely to be different in your case, but it will still probably be 172.x.y.z, and probably z will be 2. But this is not mandatory.)
So it means than inside this Docker Network, this container (one) has IP address The /16 at the end is simply here to tell you that the subnetwork is a /16. I’m not here to make a network course, but basically, it means the first 16 bits of the IP address are fixed for this network, so the last 32 - 16 = 16 bits are free to be assigned, so it means you can connect 2^16 = 65,536 containers to this network. Anyway, this is out of the scope of this article, feel free to document on this yourself.

Let’s not diverge: we just learned that our container one had IP address inside the Docker Network.

To complete the loop, we will double check the container’s IP address. Do you have an idea where we could check this?

As usually, take a break and try to answer this. I’ll be waiting (don’t cheat: it’s very important that you take some time to make your brain think about this).

Don’t be tempted to fast forward! Make you brain tick and find that answer!

Okay so you should have found out now, or at least you should have tried. (If you did not find, it’s not a big problem, but it probably means you have read too fast and did not understand every step perfectly, so scroll back up, take some sugar and go at it again).

The second place where we are going to check for the container’s IP address, is by inspecting its network interfaces, like we just did earlier. As we have seen just a couple of screens up is that our container gained a new network interface, so now we’re just going to check the IP address of the network interfaces, and it should match what’s inside the docker network inspect outout.

Here is it:

$> docker exec one ip addr show                                                                   ~
1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state UNKNOWN group default
    link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
    inet scope host lo
       valid_lft forever preferred_lft forever
14: eth0@if15: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP group default
    link/ether 02:42:ac:11:00:02 brd ff:ff:ff:ff:ff:ff
    inet scope global eth0
       valid_lft forever preferred_lft forever
20: eth1@if21: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP group default
    link/ether 02:42:ac:13:00:02 brd ff:ff:ff:ff:ff:ff
    inet scope global eth1
       valid_lft forever preferred_lft forever

Annnnnd the loop is closed: we can see that the IP address, as seem from inside the container is indeed, the same docker network inspect gives us.

Now What?

Yeah, granted, we’ve seen quite a lot, but it still lack the magical effect, with smoke, pyrotechnic effects and all that.
We’ll fix this now.

Why isn’t it all magical right now? It’s because what we have done, albeit cool, is rather limited: we have just connected a container (one) to a Docker Network (CustomNetwork).

So, basically now we’re here:

Fig 3: Container one is now inside (connected to) Docker Network CustomNetwork. Containers two and three are not.
Fig 3: Container “one” is now inside (connected to) Docker Network “CustomNetwork”. Containers “two” and “three” are not.

So this is cool because we learned a lot, but it’s limited, because it’s like having only one computer in a network: you’re not going to have much fun.

The fun part comes now: we’re going to connect another container (two), inside the same Docker Network. And that will be something!

Let’s do it right now: we’ve just seen how to do it, and it’s really fast & simple: docker network connect CustomNetwork two.

Done. See? That was quick, right?

I told you that once you undertood Docker Networks, it would be very fast to use, and very powerful.

So what changed? We’re now in this situation:

Fig 4: Container two joined one in the Docker Network. Three is still outside.
Fig 4: Container “two” joined “one” in the Docker Network. “Three” is still outside.

We can make a quick check by looking at the Containers’s section of the docker network inspect command:

$> docker network inspect CustomNetwork
"Containers": {
            "0be487e8b021b720dd657496c1ae92450c28de8ef9f1b16f06c72fe057ea567e": {
                "Name": "one",
                "EndpointID": "b090281848529d2d472f5e73350edf713d02e0480d33f006d21806ce0dbba8f3",
                "MacAddress": "02:42:ac:13:00:02",
                "IPv4Address": "",
                "IPv6Address": ""
            "8827c17a87eef24cc165e79f50df696cabcb0ce19028c41848a428010934016a": {
                "Name": "two",
                "EndpointID": "0bca67b6796de28a5666beb647e199b81b88ecbf02f6826b5f58ba99662f0a9a",
                "MacAddress": "02:42:ac:13:00:03",
                "IPv4Address": "",
                "IPv6Address": ""

Unsurprisingly, we now have one and two inside this Docker Network, and you even see that two got IP

It’s time for the smoke and confetti, part 1!

Let’s enter container one: docker exec -it one bash.

And now…. we’re going to reach to container two, remember how we just learned its IP address inside the Docker Network? Then let’s use it!

$> ping -c 4
PING ( 56 data bytes
64 bytes from icmp_seq=0 ttl=64 time=0.063 ms
64 bytes from icmp_seq=1 ttl=64 time=0.054 ms
64 bytes from icmp_seq=2 ttl=64 time=0.056 ms
64 bytes from icmp_seq=3 ttl=64 time=0.055 ms
--- ping statistics ---
4 packets transmitted, 4 packets received, 0% packet loss
round-trip min/avg/max/stddev = 0.054/0.057/0.063/0.000 ms

Hey! Did you see that?! We have made our first contact between two containers!

Let’s double-confirm that they can actually communicate, we will use netcat to emulate a client / server between the two.

Let’s create the server in two:

$> docker exec -it two bash
root@8827c17a87ee:/# netcat -l -p 1089

Let’s use the client in one and try to connect to two’s server:

$> docker exec -it one bash
root@0be487e8b021:/# netcat -v 1089
two.CustomNetwork [] 1089 (?) open

So the connection is possible and established (“open”), and you can try and write anything then press <Enter> ; it will show up on the server, in two. So communication is established!

Let’s confirm the awesomeness of this by disconnecting container two from the Docker Network: docker network disconnect CustomNetwork two. Oh and by the way it shows you the command used to disconnect a container from a Docker Network (as usual, the order is first the network name (or id) and then the container’s).

A quick look at docker network inspect CustomNetwork confirms that two is not part of the Docker Network anymore:

$> docker network inspect CustomNetwork

"Containers": {
            "0be487e8b021b720dd657496c1ae92450c28de8ef9f1b16f06c72fe057ea567e": {
                "Name": "one",
                "EndpointID": "476b9ea0975e728b862874457bd42f2b702687f8387c9eacdb0c73ea4e3b79ce",
                "MacAddress": "02:42:ac:13:00:02",
                "IPv4Address": "",
                "IPv6Address": ""

And now, if we re-try our ping command again (from inside container one):

$> ping -c 4
PING ( 56 data bytes
92 bytes from 2deaab87b3f9 ( Destination Host Unreachable
92 bytes from 2deaab87b3f9 ( Destination Host Unreachable
92 bytes from 2deaab87b3f9 ( Destination Host Unreachable
92 bytes from 2deaab87b3f9 ( Destination Host Unreachable
--- ping statistics ---
4 packets transmitted, 0 packets received, 100% packet loss

It just doesn’t work anymore.

We have to take a tiny break here and look back at what we have accomplished. Because I was not kidding with the smoke and the confetti: what we have done is really powerful.

  • We started with three separated containers: one, two and three.
  • We created a Docker Network CustomNetwork
  • Then we connected containers one and two inside this Docker Network without changing anything else.
  • And thanks to this, we were able to make one and two communicate with each other, while still being isolated from the rest of the world (noticed how three was not involved here?)

This is powerful: with that, imagine what you can do: you can connect containers running your webapp with a container running nginx, with a container running postgresql, with a container running redis, etc.

In other words, we have opened one of the last doors to building complex stacks. (In these articles, until I mention otherwise, by “stack” I mean a group of container, possibly interacting with each others through Docker Network(s), with some Docker Volumes).

I’d like you to take a small break (not one of those long break, where I show a cup of coffee icon, just a step back to appreciate the complexity of what we are doing) and enjoy (also make sure you understood what we have just done because we are going to build upon this).

This may look simple to you right now, but it’s:

  1. because I (hopefully) explain it clearly (at least that’s what I’m trying to do)
  2. because the Docker developers made a wonderful job of giving up a very nice API, simple to use ; they abstracted away all the complex notions, etc.

Now that you have appreciated this, let’s see how modular, quick and easy this setup is. Remember our friend three which was left alone up to now?
Well it’s time to connect it to our network and see that Docker Networks is much more powerful than the old & deprecated --link feature (for those who used it ; for the others, forget this last sentence).

Let’s connect two and three in the Docker Network: docker network connect CustomNetwork three and docker network connect CustomNetwork two.

A quick check:

$> docker network inspect CustomNetwork
"Containers": {
            "0be487e8b021b720dd657496c1ae92450c28de8ef9f1b16f06c72fe057ea567e": {
                "Name": "one",
                "EndpointID": "476b9ea0975e728b862874457bd42f2b702687f8387c9eacdb0c73ea4e3b79ce",
                "MacAddress": "02:42:ac:13:00:02",
                "IPv4Address": "",
                "IPv6Address": ""
            "11afdfe8e7aadb4b5acd13d434a2ca0ac118f603358e5e1aacb31ad6ed096b72": {
                "Name": "three",
                "EndpointID": "8b3c8caf6a421b3512fc1ad89a4629171a5fd093a2724004bee345084e26e5a8",
                "MacAddress": "02:42:ac:13:00:03",
                "IPv4Address": "",
                "IPv6Address": ""
            "8827c17a87eef24cc165e79f50df696cabcb0ce19028c41848a428010934016a": {
                "Name": "two",
                "EndpointID": "ee9ee79d8c1a96e359700252efb66f89ebff3faf5f71afa4becee69e4f8d0254",
                "MacAddress": "02:42:ac:13:00:04",
                "IPv4Address": "",
                "IPv6Address": ""

So we do have all three containers in our Docker Network. And you can enter any of them and ping any of the other’s IP address and that will work. Similarly you can use netcat to create a server on any of these containers and use netcat as a client from any other container to contact to it and check that data exchange is really possible.
Disconnect any of the container from the network, and it will simply cease to be reachable by the others.

Isn’t that all wondeful?

What? What do I hear?
Yes I can see you in the back… what’s the problem?

You’ve noticed that container two has now IP address and not like it had before we disconnected / reconnected it?
So what, you want a medal?

Just kidding! Very nice catch actually.

As you may have guessed, Docker Networks tend to address IP addresses in an increasing number, in the order containers are connected. This is a trend, but please you should not rely on this because this is not documented anywhere, so the docker dev teams are free to change this without prior notice. Actually relying on this for you stack is a guarantee for trouble.

While it can be fine for “static” configurations, it might become a very big problems for applications that are more dynamic: what if containers can be created on the fly and join the Docker Network, then maybe become disconnected, etc.?

This is a problem.

So we need a mechanism to discover a container’s IP address inside the network, etc. This seems complex.

Actually, as it’s the case for a lot of things, the Docker devs have thought about this, and it’s time for smoke and confetti part 2!

I’m going to say (or rather write) a sentence that will sound complex and might make you think “okay, I reached my limit with Docker, let’s give up”, but please don’t, you’ll see how ridiculously easy this actually is. Okay here it is:

In each custom Docker Network, Docker provides an embedded, internal DNS server that allows any container to resolve the other container’s IP addresses based on their container’s name

Nobody moves! Stay here!

Seriously, this sentence above means exactly this: “inside a Docker Network, you can reach a container by its name, rather by its IP”.

Ah?! Easy or not?! Didn’t I tell you?

So picture this: we have been ping-ing IP addresses like crazy for half an hour, using docker network inspect to find the container’s IP address while there was a nice, and reliable feature that was awaiting!

It’s time we tried this, no?

Make sure your three containers are connected to CustomNetwork, enter container two and try pinging two and three:

$> ping one -c 4
PING one ( 56 data bytes
64 bytes from icmp_seq=0 ttl=64 time=0.070 ms
64 bytes from icmp_seq=1 ttl=64 time=0.057 ms
64 bytes from icmp_seq=2 ttl=64 time=0.058 ms
64 bytes from icmp_seq=3 ttl=64 time=0.056 ms
--- one ping statistics ---
4 packets transmitted, 4 packets received, 0% packet loss
round-trip min/avg/max/stddev = 0.056/0.060/0.070/0.000 ms

$> ping three -c 4
PING three ( 56 data bytes
64 bytes from icmp_seq=0 ttl=64 time=0.085 ms
64 bytes from icmp_seq=1 ttl=64 time=0.057 ms
64 bytes from icmp_seq=2 ttl=64 time=0.058 ms
64 bytes from icmp_seq=3 ttl=64 time=0.056 ms
--- three ping statistics ---
4 packets transmitted, 4 packets received, 0% packet loss
round-trip min/avg/max/stddev = 0.056/0.064/0.085/0.000 ms

For consistency’s sake, we will create a server with netcat on container three and try to connect from two:

$> docker exec -it three bash
root@11afdfe8e7aa:/# netcat -l -p 2000

Yes I changed the port so that you can see for yourself that everything works for real.

$> docker exec -it two bash
17a87ee:/# netcat -v three 2000
three.CustomNetwork three [] 2000 (?) open

And then I’ll let you write anything in the client and press <Enter> and see that it gets written back on the server.

Notice here the big difference: in the client, we did not use three’s IP address, but three’s name: three.
In this case, we don’t even know (care?) what three’s IP is: Docker resolves it for us!

Ah ah! Seriously, look me into the eyes and tell me: how awesome is that? As you can see here, when we pinged one, it resolved to, and when we pinged three, it resolved to

Let’s make sure it works: disconnect one and three: docker network disconnect CustomNetwork one && docker network disconnect CustomNetwork three and reconnect them, in the other reverse order (to change their IP address); docker network connect CustomNetwork three && docker network connect CustomNetwork one.

Let’s check that their IP addresses have been changed:

$> docker nework inspect CustomNetwork
"Containers": {
            "0be487e8b021b720dd657496c1ae92450c28de8ef9f1b16f06c72fe057ea567e": {
                "Name": "one",
                "EndpointID": "da6c82447d34726b9e08977ea13a31411aadb238a08cd7b0fec7b293c9007c02",
                "MacAddress": "02:42:ac:13:00:03",
                "IPv4Address": "",
                "IPv6Address": ""
            "11afdfe8e7aadb4b5acd13d434a2ca0ac118f603358e5e1aacb31ad6ed096b72": {
                "Name": "three",
                "EndpointID": "95c3acb0ebecb52de368bea7ecb0a57f45ed0ec02606f00a7629765671fff77e",
                "MacAddress": "02:42:ac:13:00:02",
                "IPv4Address": "",
                "IPv6Address": ""
            "8827c17a87eef24cc165e79f50df696cabcb0ce19028c41848a428010934016a": {
                "Name": "two",
                "EndpointID": "ee9ee79d8c1a96e359700252efb66f89ebff3faf5f71afa4becee69e4f8d0254",
                "MacAddress": "02:42:ac:13:00:04",
                "IPv4Address": "",
                "IPv6Address": ""

Yup. So enter two again, and ping the container again:

root@77e3e44ee3cc:/# ping -c 4 one  
PING one ( 56 data bytes
64 bytes from icmp_seq=0 ttl=64 time=0.048 ms
64 bytes from icmp_seq=1 ttl=64 time=0.052 ms
64 bytes from icmp_seq=2 ttl=64 time=0.053 ms
64 bytes from icmp_seq=3 ttl=64 time=0.053 ms
--- one ping statistics ---
4 packets transmitted, 4 packets received, 0% packet loss
round-trip min/avg/max/stddev = 0.048/0.051/0.053/0.000 ms

root@77e3e44ee3cc:/# ping -c 4 three
PING three ( 56 data bytes
64 bytes from icmp_seq=0 ttl=64 time=0.052 ms
64 bytes from icmp_seq=1 ttl=64 time=0.058 ms
64 bytes from icmp_seq=2 ttl=64 time=0.057 ms
64 bytes from icmp_seq=3 ttl=64 time=0.055 ms
--- three ping statistics ---
4 packets transmitted, 4 packets received, 0% packet loss
round-trip min/avg/max/stddev = 0.052/0.056/0.058/0.000 ms

Annnnd it works! Wonderful.
So we now have a very handy tool: you don’t ever have to know / find a container’s IP address, because you can simply use its name. Isnt’ that awesome? Seriously?

Let’s disconnect three now with docker network disconnect CustomNetwork three and try to ping it again:

root@77e3e44ee3cc:/# ping -c 4 three
ping: unknown host

As soon as the container disconnects from the Docker Network, it is not reachable anymore. No need to realod any configuration or anything: the container left the network, it is not there anymore. Period. Awesome. Confetti.

A Quick Gotcha

Before we make a nice little recap of all the cool stuff we have just seen, I wanted to show you a small gotcha without which, I am sure, some of you were bound to bang their head against the wall until one of the two breaks (and from experience, it’s often the head!). The very last thing we have just discovered is that inside a Docker Network, you don’t need to know a container’s IP address to talk to it, just its name.
This is why we can do ping three and it automatically resolves to the correct IP address.

This is all good and all, but just know that it only works with explicitly-named containers. In other words, it doesn’t work with containers that are named automatically by docker, let’s see this in effect:

$> docker run --rm -itd net-test bash

$> docker ps --format "{{.Names}}"

Above we have created a new container without naming it, then we have used docker ps to list container (and a nice little Go template to only display the names, because docker ps is very verbose.) Obviously, your container will most likely have another name, just replace accordingly (it’d be fun if just for this example your container is also named like this!).

Then it’s time to connect this container to our Docker Network with docker network connect CustomNetwork peaceful_lovelace (tip: you can use Tab-compltetion for the commands and the container names).

Then let’s enter container one for instance and ping this new container:

$> docker exec -it  one bash

root@2deaab87b3f9:/# ping -c 4 peaceful_lovelace
ping: unknown host

On ma machine it takes about 30 seconds, but it eventually times out.

All of this to say that this is weird, but it’s the rule (at least, for now): this automatic container name resolution only works with explicitly-named containers. Just remember it before thinking your Docker Networks has is broken.

There is one last workaround to this rule: --network-alias.
This is an option to docker run which, as its name implies, gives a container an alias, for network communication. What this means is that whatever the name of the container, it can also be resolved on the network by one of its network aliases.

So if you create a container without the --name but with a --network-alias my-alias then any container on the network can reach this one by using my-alias. It acts as the name of the container.

This is useful if, for instance, you want a container to be reachable from several names. You can have a container that is named test-postgres-v4 so it is reachable in the Docker Network with test-postgres-v4. But you can also give it --network-alias db and it will be reachable with db. This is especially useful if you are testing a new configuration (in a new container): simply give your new container an network alias that corresponds to the name the other containers used to contact the old (legacy) container, and it will be transparent.

Time For a Recap!

That was a long one, right?! I’m not talking about the article: we’re about halfway through it! I’m talking about this last part. This was quite the beast.

We have learned a lot, so it’s time to make a recap, because this stuff is very important (seriously, Docker Networks is absolutely necessary if you want to make complex stacks and want to become a tiny bit advanced with Docker).

So what have we learned?

At the start of this article, we knew quite a lot about containers, and especially that they are isolated: containers are processes on the host, that live in different namespaces and they cannot interact with each other. Which is very good because that’s what we are looking for with Docker, but can also be a problem when you start building stacks that require communication between two (or more) containers.

One obvious way to “solve” this problem would have been to run all the applications that need to communicate in the same container. But that would be going against Docker’s principles if the processes are not intimately linked, and this is obviously not something we can use for large and complex stacks.

We briefly talked about a legacy way, --link which is a rough way to directly connect two containers. It was the only solution at some point in time, but now it should be banned.

Then we saw the solution: Docker Networks. They are as easy to create as containers (docker network create and docker network rm) and they provide networking features for our containers. A Docker Network is totally separate entity from a container, which means that you can “hot-connect” an existing, running container inside a Docker Network (i.e. connect the container without stoppping or destroying it).
When a container gets connected to a Docker Network, it gains a new network interface, because as far as this container is concerned, it as if its host system just had a new network interface installed and joined a new local network (like a LAN).
The container thus gains a new IP address (its IP address inside the Docker Network) that can be used to communicate with it, by the other containers in the same Docker Network.

The very handy thing about Docker Networks is that they are a separate entity (from the containers). It’s something that you create an dhandle separately. And you can, at some point (i.e. not necessarily at the time you run your container) you can decide to connect your container to a Docker Network (it then becomes reachable by the other containers inside this Docker Network), and whenever you want, you can disconnect it from the Docker Network (thus becoming unreachable).

A nice little feature is the embedded DNS server that allows containers to talk to each other through their container’s name (and/or their network aliases) instead of going through the process of discovering the target container’s IP address. Remember, though, that it works only with explicitly-named containers (or containers that have network alias(es)).

Slightly More Advanced Use-Cases

A Nice Shortcut

Now that all of this is slowly sinking in, let me show you one small option that should save you some time. Until now we have instantiated a container with docker run, we usually gave it a name with the --name option, and then we connected the container to our Docker Network with docker network connect <network> <container>.

This is fine and works. But just for the sake of mentionning it, know that it is possible to directly connect the container to a Docker Network thanks to the docker run command. This way you can do it all in one command.
This doesn’t change anything: the behavior is exactly the same, and if you find this confusing, feel free to forget about this. But if you’re interested, the option to add is --net.

So if you want to connect your container to a Docker Network named CustomNetwork directly on container creation, you can do something like: docker run -it --name my-container-name --net CustomNetwork <image-name> and voilà: when your container is created, it is already connected to CustomNetwork.
Obviously if at some point you want to disconnect it from the Docker Network, docker network disconnect is your friend.

Anyway, that was the easy part, I have something else to tell and show you, are you ready? This is going to be a good one.

Add Some More Fun!

I don’t know of a better way to tell you, so I’ll just blurt it out:

You can connect a container to several Docker Networks at the same time!

Yes, you read that correctly. Remember how we said that when we connect a container to a Docker Network, it gains a new network interface, because for the processes running inside the container, it is as if the machine had a new network card and just joined a new LAN network? Well what prevents you from installing a second network card and connect to another LAN then?
Nothing. And this is also possible with Docker.

Until now, we had situations like this:

Fig 5: Simple configuration with several containers in the same Docker Network
Fig 5: Simple configuration with several containers in the same Docker Network

Now we will do something like this:

Fig 6: Complex configuration where Container C is connected to two Docker Networks simultaneously
Fig 6: Complex configuration where “Container C” is connected to two Docker Networks simultaneously

And we will see that it works perfectly as intended.

Let’s create two Docker Networks:

$> docker network create Network1

$> docker network create Network2

And let’s now create three containers:

$> docker run -itd --name ContainerA net-test

$> docker run -itd --name ContainerB net-test

$> docker run -itd --name ContainerB net-test

Now, we will connect ContainerA to Network1 and ContainerB to Network2, like we already did several times:

$> docker network connect Network1 ContainerA
$> docker network connect Network2 ContainerB

Nothing new here. The fun part comes from ContainerC. We want to connect it to both Network1 and Network2. Well, let’s do this. First, let’s connect it to Network1: docker network connect Network1 ContainerC.

We can see that now both ContainerA and ContainerC are connected to Network1:

$> docker network inspect Network1
"Containers": {
            "0407e1bad71112464bfc1d3cafa284549c2504400482a5103998352b5eb6aa07": {
                "Name": "ContainerA",
                "EndpointID": "34ba3ffd2e059b4e9135aaf818b2946d33cf02fcc307d4fa500c88c42baf39cd",
                "MacAddress": "02:42:ac:14:00:02",
                "IPv4Address": "",
                "IPv6Address": ""
            "f18c6c868e9a3b283b59d0b41df8696eadf9bebfd3da405117bd4313b48ccf67": {
                "Name": "ContainerC",
                "EndpointID": "c75608e678849752975e1f0a64ff3b229f6964ac162fe84aba8854d285eef876",
                "MacAddress": "02:42:ac:14:00:03",
                "IPv4Address": "",
                "IPv6Address": ""

And they can reach each other:

$> docker exec ContainerA ping -c 3 ContainerC
PING ContainerC ( 56 data bytes
64 bytes from icmp_seq=0 ttl=64 time=0.073 ms
64 bytes from icmp_seq=1 ttl=64 time=0.058 ms
64 bytes from icmp_seq=2 ttl=64 time=0.059 ms
--- ContainerC ping statistics ---
3 packets transmitted, 3 packets received, 0% packet loss
round-trip min/avg/max/stddev = 0.058/0.063/0.073/0.000 ms

(Here, instead of first “entering” ContainerA and then run a ping command to ping ContainerC, I use docker exec to execute a command inside the container, this command being ping -c 3 ContainerC. This saves us a roundtrip of “entering the container”, making our call, exiting the container. But it’s equivalent. Also, I will let you test that it still works with netcat.)

Just to be extra sure, let’s check that ContainerC is not in Network2:

$> docker network inspect Network2
"Containers": {
            "c12bfd2e1842404f1c66b031b8d79225375740aa6daed0123287c5f355804417": {
                "Name": "ContainerB",
                "EndpointID": "ca2a694e9acc36ffd6ab0bb71a7a4830c4c9f26a40411cd7f0a4523424b4d58a",
                "MacAddress": "02:42:ac:15:00:02",
                "IPv4Address": "",
                "IPv6Address": ""

Perfect. All is normal at this point.

Last check before doing anything else (and to really follow what we are doing), let’s check ContainerC’s network interfaces:

$> docker exec ContainerC ip addr show
1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state UNKNOWN group default
    link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
    inet scope host lo
       valid_lft forever preferred_lft forever
12: eth0@if13: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP group default
    link/ether 02:42:ac:11:00:04 brd ff:ff:ff:ff:ff:ff
    inet scope global eth0
       valid_lft forever preferred_lft forever
16: eth1@if17: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP group default
    link/ether 02:42:ac:14:00:03 brd ff:ff:ff:ff:ff:ff
    inet scope global eth1
       valid_lft forever preferred_lft forever

So we see the usual lo loopback interface and the usual eth0 interface that I briefly mentionned (remember: for now consider that when you create a container, it starts with two network interfaces: lo and eth0; I promise that the real explanation is coming soon).
And we can see eth1 which corresponds to the network interface this container has in the Docker Network Network1 (check the IP addresses, you’ll see they match!)

Okay, perfect! It’s time to connect ContainerC to Network2 now: docker network connect Network2 ContainerC.

Same deal: let’s check everything. First, the network:

$> docker network inspect Network2
"Containers": {
            "c12bfd2e1842404f1c66b031b8d79225375740aa6daed0123287c5f355804417": {
                "Name": "ContainerB",
                "EndpointID": "ca2a694e9acc36ffd6ab0bb71a7a4830c4c9f26a40411cd7f0a4523424b4d58a",
                "MacAddress": "02:42:ac:15:00:02",
                "IPv4Address": "",
                "IPv6Address": ""
            "f18c6c868e9a3b283b59d0b41df8696eadf9bebfd3da405117bd4313b48ccf67": {
                "Name": "ContainerC",
                "EndpointID": "572f8e062bf7ed9b0c9c9679c18076376c3e0b2c57567cbcf4e27b2cd08ffcc7",
                "MacAddress": "02:42:ac:15:00:03",
                "IPv4Address": "",
                "IPv6Address": ""

Hey! so it seems it works the same: we previously had only ContainerB, but now, we also have ContainerC.

Let’s check that ContainerB and ContainerC can talk to each other:

$> docker exec ContainerB ping -c 3 ContainerC
PING ContainerC ( 56 data bytes
64 bytes from icmp_seq=0 ttl=64 time=0.074 ms
64 bytes from icmp_seq=1 ttl=64 time=0.053 ms
64 bytes from icmp_seq=2 ttl=64 time=0.056 ms
--- ContainerC ping statistics ---
3 packets transmitted, 3 packets received, 0% packet loss
round-trip min/avg/max/stddev = 0.053/0.061/0.074/0.000 ms

Of course that works: they are part of the same Docker Network!

If everything is coherent, ContainerC must have gained a new network interface: the one connect it to Network2:

$> docker exec ContainerC ip addr show
1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state UNKNOWN group default
    link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
    inet scope host lo
       valid_lft forever preferred_lft forever
12: eth0@if13: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP group default
    link/ether 02:42:ac:11:00:04 brd ff:ff:ff:ff:ff:ff
    inet scope global eth0
       valid_lft forever preferred_lft forever
16: eth1@if17: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP group default
    link/ether 02:42:ac:14:00:03 brd ff:ff:ff:ff:ff:ff
    inet scope global eth1
       valid_lft forever preferred_lft forever
20: eth2@if21: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP group default
    link/ether 02:42:ac:15:00:03 brd ff:ff:ff:ff:ff:ff
    inet scope global eth2
       valid_lft forever preferred_lft forever

Yes sir! It did alright!
As you can see, ContainerC now has eth2, with the IP that matches the one in the output of when we inspected the Docker Network.

Everything seems fine, we should be in the situation depicted in Fig. 6 now, two checks are left to be made:

  1. check that ContainerC, being in both Network1 and Network2 can still communicate with ContainerA (via Network1) and ContainerB (via Network2).
  2. check that ContainerA and ContainerB cannot communicate, because they are not part of the same Docker Network, which would confirm that just because they both know a container (ContainerC) than is in both Networks, nothing is leaked out.

Okay, check #1:

$> docker exec ContainerC ping -c 3 ContainerA
PING ContainerA ( 56 data bytes
64 bytes from icmp_seq=0 ttl=64 time=0.051 ms
64 bytes from icmp_seq=1 ttl=64 time=0.056 ms
64 bytes from icmp_seq=2 ttl=64 time=0.068 ms
--- ContainerA ping statistics ---
3 packets transmitted, 3 packets received, 0% packet loss
round-trip min/avg/max/stddev = 0.051/0.058/0.068/0.000 ms

$> docker exec ContainerC ping -c 3 ContainerB
PING ContainerB ( 56 data bytes
64 bytes from icmp_seq=0 ttl=64 time=0.063 ms
64 bytes from icmp_seq=1 ttl=64 time=0.052 ms
64 bytes from icmp_seq=2 ttl=64 time=0.069 ms
--- ContainerB ping statistics ---
3 packets transmitted, 3 packets received, 0% packet loss
round-trip min/avg/max/stddev = 0.052/0.061/0.069/0.000 ms

So it works! ContainerC can talk to both. Notice how the subnet changes (172.21.x.y) vs (172.20.x.y): the container switches network interface to switch network.
Note that I did not say “note how the IP address changes”, that is totally obvious (we’re talking to a difference container, so of course the IP address changes) ; I said that the subnet changed.

Now for check #2:

$> docker exec ContainerA ping -c 3 ContainerB
ping: unknown host

Perfect! As predicted (and desired!), ContainerA and ContainerB do not see each other, they can’t talk to each other. Isolation is kept. Seriously, this is awesome.

What About This “default” eth0 Network Interface?

Hum… so you remember this, right?
And you want me to explain it now? Well… I guess I have no choice, then.

I said I would explain later, so I might as well do it now!

Okay for those who have forgotten, if you recall correctly, when we created containers, it already had a network interface, usually named eth0@xxx. Then when we connected this container to a Docker Network with docker network connect <network> <container>, it gained a new network interface, usually named eth1.

So what is that eth0?

Let’s make a few experiments before I explain, so that you can see for yourself.

Let’s create a container, as usual and notice this phenomenon:

$> docker run --rm -itd --name C1 net-test bash

$> docker exec C1 ip addr show                                                                    ~

12: eth0@if13: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP group default
    link/ether 02:42:ac:11:00:02 brd ff:ff:ff:ff:ff:ff
    inet scope global eth0
       valid_lft forever preferred_lft forever

(For this section I will remove the lo interface because it is irrelevant for our example, and will help make the article not too long).

Okay, so it does have this eth0 network interface. Let’s now connect it to CustomNetwork.

$> docker network connect CustomNetwork C1

$> docker exec C1 ip addr show
12: eth0@if13: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP group default
    link/ether 02:42:ac:11:00:02 brd ff:ff:ff:ff:ff:ff
    inet scope global eth0
       valid_lft forever preferred_lft forever
14: eth1@if15: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP group default
    link/ether 02:42:ac:13:00:02 brd ff:ff:ff:ff:ff:ff
    inet scope global eth1
       valid_lft forever preferred_lft forever

Okay. Makes sense. If we disconnect it from the Docker Network, it will lose its eth1 interface.

Let’s do this again (keep this container alive for now).

docker run --rm -itd --name C2 net-test bash

$> docker exec C1 ip addr show
16: eth0@if17: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP group default
    link/ether 02:42:ac:11:00:03 brd ff:ff:ff:ff:ff:ff
    inet scope global eth0
       valid_lft forever preferred_lft forever

So as expected this C2 container also has a eth0 network interface. But don’t you notice something peculiar?

I’ll let you search for a couple of seconds.

Nothing? Let’s make one more test and see if it ticks:

$> docker run --rm -itd --name C3 --net CustomNetwork net-test bash                                  ~

$> docker exec C3 ip addr show
18: eth0@if19: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP group default
    link/ether 02:42:ac:13:00:03 brd ff:ff:ff:ff:ff:ff
    inet scope global eth0
       valid_lft forever preferred_lft forever

Okay, again this test? We know this, this is nor… wait what?

Yep! Look again at this last command, this time I used the --net CustomNetwork option to connect C3 directly. And what happened? The container has the familiar eth0 network interface, but that’s all.

I’ll let you think about this now.

Okay, now’s the time for the big reveal (note that you have everything you need to find the answer yourself, though this is not easy).

Okay so the last example is not that trivial. The eth0 that we see is not the one we’re used to. Look closely at C1, C2 and C3.
In the first two, eth0 is in the 172.17.xx.yy subnet, but in C3, it is in 172.19.xx.yy subnet, which corresponds to the subnet eth is in for containers C1 and C2.

So what is really going on here?

When you create a container without the --net option, it gets automatically connected to a default bridge network. This bridge network you can see while listing the Docker Networks:

docker network list                                                                            ~
NETWORK ID          NAME                DRIVER              SCOPE
9aedc6a27184        CustomNetwork       bridge              local
70ff2194d229        MyFirstNetwork      bridge              local
83b7a5902813        Network1            bridge              local
721f118ce184        Network2            bridge              local
e68176304a37        bridge              bridge              local   # <---- that's the one!
2683c3824f8f        host                host                local
a84e731c4387        none                null                local

And if you inspect it:

$> docker network inspect bridge
"Containers": {
            "134357a53de8a5778a3a6a405321156e345563379186c638a8373bb477f52e58": {
                "Name": "C2",
                "EndpointID": "840aa7fa22a92b38c8e2d9b3de42999cfe5fba9a8ec7b42c5f6d676dce0e6dc6",
                "MacAddress": "02:42:ac:11:00:03",
                "IPv4Address": "",
                "IPv6Address": ""
            "220c4e21c10e4a0f6b90e0015fc5f40f221437cd0675164eeb0a0e271e4e6c29": {
                "Name": "C1",
                "EndpointID": "6445d819e6edb6209fb1272abeb8c8380313a61e44d74da8d8e7514f94e84219",
                "MacAddress": "02:42:ac:11:00:02",
                "IPv4Address": "",
                "IPv6Address": ""

And here you can see that containers C1 and C2 are indeed connected to this Docker Network, even though we did not specify --net bridge.

And since they are connected to this Docker Network, they gain a network interface (eth0) which is in bridge’s subnet.

As for C3, since we created it with --net CustomNetwork, it got connected to CustomNetwork and that’s all, this is why it only has eth0 and that it’s subnet is CustomNetwork’s.

The TL;DR of this is that when you create a container without any --net option, it gets automatically connected to the default bridge network. That’s a golden rule of Docker that you should never forget.

This means that usually, there are a lot of containers in that default bridge Docker Network.

Let’s now kill two birds with one stone.

Since there are usually lots of containers in this bridge Docker Network and this can potentially lead to confusion, the embedded DNS server is not present in bridge. This means that even if you named your container with --name <container-name>, you cannot reach this container with the hostname <container-name>.
In the bridge Docker Network, you can only reach a container by its IP address.

This is made to prevent confusion. If this seems not appropriate for production, it is normal: it is also meant to discourage you to use the default bridge network in production. If you want to have a clean production system, you should always use a custom Docker Network.


Okay a tiny recap section with two nice quotes, because really, you should not forget this (for your information, of all the time I spend on #docker, I repeat the two quotes at least once a week).

There are basically two gotchas that you need to remember:

Every container not created with the --net option is automatically connected to the default bridge Docker Network


In the default bridge Docker Network, there is no embedded DNS server, which means automatic name resolution does not work

That’s all. Make sure to always use custom Docker Networks when doing serious work.

Can You Tell us More About the Other Networks?

Oh boy you are curious!
That’s good, I love curious people (maybe we can meet…? No? … Okay).

Back when I introduced the Docker Networks, when we first ran docker network ls, we saw three default Docker Networks:

  • bridge which we just explained (you know the default network to which all containers are connected by default, unless you specify a --net parameters)
  • host
  • none

Okay, so none is the easiest to understand: basically (as the name slightly implies!) it is used when you want to create a container but do not want to connect it to any network at all. In particular, creating a container with --net none will prevent this container to be automatically connected to the default bridge network.
But be warned: none is also meant to enforce a container to have no connectivity, at all. So for as long as your container is part of the none network, it cannot be connected to any other Docker Network. You first need to disconnect it from the none network to connect it to another Docker Network.
Be sure to remember this.

host is different and a little special. First of all, you cannot connect or disconnect a container to / from the host network. The only thing you can do is create a container with the --net host option. This way it will be automatically connected to the host network when it is created, and that’s all.
You cannot disconnected it later (you have to plainly destroy the container) and if you did not create the container with --net host, it not possible to add it later.

So what does it do then?

Well, when you make your container part of the host network, it means the container becomes part of the host’s network stack. It means it will share all of the host’s network interfaces. In other words, using --net host means you are not isolating your container (on the network side): it is a “normal” process on your host (from a network point of view), so it can see and access the host’s network interfaces.

To be honest, if you end up using --net host for another reason that testing & curiosity, it’s either because you’re taking a nasty, dirty shortcut and you don’t understand what you do, or you are really advanced and you know perfectly what you are doing.
Personally, I have never had to use --net host for real, there is (almost) always a better, cleaner solution.
So I included the explanation for consistency’s sake, but I really discourage you to use --net host.

Small Recap?

This was not long, so let’s make this recap short!

We have finally learned about the three default networks that come “pre-installed” when you install Docker. They are: host, none, bridge.
Note that you cannot remove these three networks: if you try docker network rm on any of them the world falls apart you’ll get insulted by Docker.

The roles of these networks are:

  • bridge: this is the default bridge network to which all containers are connected when you create them without a --net option. Note that automatic name resolution (pseudo embedded DNS server) is not working for this network. Do not use on production.
  • none: used to prevent a container to have any sort of network connection
  • host: prevent isolation of a container on the network side: it will share the host’s network stack unaltered.


So that was quite a part, wasn’t it? We have learned one of the most powerful (and one of my favorite) Docker features.
If you recall what we had before this part, we could create separate and individual containers. We learned how to create good and well-structured Dockerfiles to eventually be built and create images.
We learned how to make all of our containers totally ephemeral, i.e. we made it so that out containers could be stopped, restarted, destroyed and recreated at any time, without problems thanks to making important data persistent with Docker Volumes.

This was all good, but had a feeling of “incompleteness”: it allowed only “single-container” applications, or, worse it made you cram all your application processes inside one, monolithic container. This is bad. You should not do it anymore. For real (otherwise you’ll go to the same place in hell as those that keep using --link!).

Now, thanks to Docker Networks, we learned how to break the isolation in a controlled way between several containers: essentially, we learned how to make them communicate. And I hope you can see how many opportunities this now creates!

You can totally think about containerizing every part of your application stack: separate the database from the web server, from the intermediate redis cache, from a custom-implemented CDN too of your wish!

But believe me, this is only half the fun (actually, this is even one third) of the fun!

What do I mean by that?

I mean that, network-related, we still have two awesome features to discover in Docker: overlay networks and the Swarm Mode. You might have heard these names before, you might even have tried to use them a little. If yes, then you should be excited, because we’re going to talk about these. If no, well.. you should be excited, because it’s amazing (and we’re still going to talk about them!).

The second part of this Part V article will be about overlay networks, because they are very very close to the bridge-based Docker Networks we have been studying up to now. As for the Swarm Mode, since this is a huge topic in itself, this is going to be covered in another Part.

So now we are going to be talking about overlay networks, but before that, you really need to take a coffee break. And as usual, I mean it: a lot of the features we are going to be talking about have the same name, and deal with the same concept. This is the perfect way to get things mixed up. You think you understand Docker Networks, you think you mastered them bcause you’ve just read about them, and you end up mixing everything.
So as usual, here is the cup-of-coffee image, take a break, the article will still be there when you return!

Multi-Host Networking to Control the World!

And here we are! This section on overlay networks should complete your knowledge about Docker Networks, and let me tell you that once you’ve mastered this, there is very little you cannot do with Docker.
At the end of this section (and this Part), you will have all the fundamental notions needed to make anything complex in Docker.

Excited? Let’s read on, then!


Just before I say anything else, just let me make something clear. If you look at an overview of this article, you will see that it is quite long, and the part about overlay networks comes at the end, and is relatively short, compared to the rest of the article. So I would understand if you went and think this is just a detail, it’s trivial or not that important. You’d be wrong.
Overlay networks are even more powerful that bridge-based networks (the Docker Networks we have been talking about since the beginning of this article). But the Docker developers made their homework and made the API and most of the concepts work transparently. In orther words, there is a reason I’m talking about overlay networks after bridge-based networks: it’s because it re-uses all the notions we saw there.
So really, don’t think overlay networks are trivial: they simply have a lots in common with bridge-based networks, they simply have additional features, so you need to read the whole Part above to understand overlay networks.

That being said, let’s proceed!


So with everything we have seen so far, we can have complex stacks: we can design our own images in Dockerfiles, perfectly suited to our needs, we can make data persistent, and thanks to Docker Networks, we can separate all logical aspects of our application into separate containers, make them communicate, and all is (almost) good in the universe.

Yes but, now if we scale up a little, we might encounted a situation. What if you want to deploy your setup in production and use dedicated servers to deal with the different aspects of the system?
For instance, you can have a dedicated server to handle all incoming traffic, with a tailor-made nginx server running on it. If you want performance at this level, there are some choices to be made and you might end up with a server that is perfectly suited to route traffic and make load balancing, but not for handling lots of SQL requests.

So here you are, buying a second server that you fine-tune to be as performant as possible with the task of dealing with SQL requests.

Or you might forget what I just said and suppose that you prefer to keep the data in-house and use a home-hosted server for the database, but you’re okay with renting / buying a server on a provider to handle the routing.

Or you might even craft other scenarii. The point is: it is very likely that some of you find themselves in a situation where the several components of your application stack do not live on the same machine ; that is the point we will address today.

In the first parts of this article, we learned how to separate our logical units (processes), for instance an nginx and a postgresql server, run them in their own container, put them in a Docker Network and have them communicate like this. But our problem now is: how do we do when the nginx container and the postgresql containers are not on the same computer?!

It’s not an unsolvable problem: we might totally imagine dockerizing the two (or more) services, each on one host (“host” is the Docker term we use to talk about the “machine” / “server” / “computer” on which Docker is installed), bind the ports (we know how to do this already) and provided the containers have network accessibility, they can talk to the other remote hosts like a normal application would do.
So you are not out of options, obviously. It works. I mean, this is what people use all the time: making two (or more) computers communicate is not rocket science.

Obviously, it’s your responsibility to provide either an IP address or a domain name for the dockerized applications to be able to talk one to another, it’s also your responsibility to ensure the communication is encrypted, etc. But that is manageable.

But if we take a step back, we can see that our whole setup is very similar to the ones we have been building in the first parts of this article, where we dockerized individual components on the same host, and that would be very nice and handy if we could somehow use the same mechanisms (Docker Networks), but on multi host (several different machines).

And yes we can! Let’s see how!

Conceptually Identical, Implementation (almost) Identical

As you must have understood from this little introduction, the docker developers have thought about this use-case, and provided us with a way. And the very nice part is that you will not need to learn much.. yes you guessed it, we will still use Docker Networks.
So everything we have just seen together still applies (and this is why you should read the beginning of this article, otherwise it will sound like gibberish).

So what’s the difference then?

Do you remember when you used docker network ls? Try it now, there is a column, DRIVER which we have talked about before. It was one of three values: host, bridge or null.

Actually, there is a fourth value: overlay, which is part of the magic behind what we are about to see.

What to Expect?

Since this article is already very long, and you must be tired (although not that much if you followed the breaks I recommended), it’s best if I don’t discourage you heads on. The really good thing here is that you will have almost nothing new to learn. Really!

Everything we saw still applies, and from a developer’s point of view, it works almost exactly the same! Considering the time we took to study this, the length of the article, I personally think this is a great example of high-level API and abstraction, and I’d like to thank the docker developers for this.

I deal with code all day long and spend a great deal of cringing on poorly-written code, well the docker devs have made a good job.

Anyway, back to business. So as I was saying, you won’t need to learn much, the details, the magic is handled at lower levels. And this is great.

To be honest, I initially planned on explaining how Docker Networks worked internally in details, but the article would be much too long. So I have removed it from now, and will make a dedicated article later if you are interested ().

In order to keep everything short (I can already see two or three people sleeping in the back of the class), I will only give you the bare basics (and you know I hate this…), this should be enough for you to go read some external documentation about the topics.

So How Does This Work?

Earlier in the article when I introduced the Docker Networks, especially the bridge-based ones, I talked about virtual network interfaces (vethXXXX) that were seen from inside the container. The inner working of the Docker Networks we have seen up to now (the bridge-based ones) is elegantly simple: this relies on iptables.

For those who don’t know iptables, this is a Unix utility used to control the routing & filtering of IP packets. “This is a firewall” to put it simply. iptables work with rules and targets. Very roughly, you configure it to say “packets of this protocol from this source are ALLOWED / DENIED to reach this destination”. Another rule is the FORWARD rule, which basically reroutes packets.

This is what Docker uses internally: when it want to make several containers communicate, it routes packets between the virtual network interfaces. This is elegant, simple and very efficient.

But obviously, this works only locally, i.e. on the same host, because you need to know and access the network interfaces in order to do this. So although the API (the high-level part, the docker network commands) is similar, the internals are different.

Now suppose you have another computer (or “host”), say a remote server that has docker installed on it, and you have a container on it. Now you have your own host, with another container running and you’d like to estalish a Docker Network between the two. Docker needs to communicate with the remote computer and exchange information to synchronize the two docker daemons.

This is not handled internally. Why? Because when you take a step back from this setup, it’s not really docker-dependant. The problem reduces to a data synchronization problem, between two (or more) computers. And this is not a new problem, that was “solved” before. Not much data needs to be stored, basically, what container is on which networks, the IP address of the host and the internal IP adress op the Network.

In order to make this possible, Docker uses a (distributed) Key-Value store. A Key-Value store (KV-store in short) is a very simple array-like “database”. Values are relatively simple, and they are indexed by a key. There exists several popular and well-functionning KV-stores, and thus Docker leveraged them.

So the TL;DR of all of this is: in order to set up multi-host (or “overlay”) Docker Networks, you install and setup a KV-store, configure your Docker Daemon to interface with it and you’re done. How awesome is that?.

I have just described the structure of the rest of the article, so let’s go!

Hey! I don’t Have a Server on Hands!

Okay okay, I know.
So if you don’t have yet a remote server on which you can install Docker, so that you can try multi-host network, fear not, we have a workaround.
For such cases, Docker provides us with a new tool, called “docker-machine”. “docker-machine” can do several things, one of which is to create optimized, small-size Virtual Machines in which Docker is installed.

I’m going to show you how to setup a cluster of docker-machines to emulate some remote servers. In the rest of the article I’m going to try and show both ways: with a server and with docker-machines.

I just want to clarify things, in order to prevent confusion: in our case, docker-machine’s only purpose is to create separated virtual machines that we will use as if they were remote servers. That’s all.
If you do have a remote server on hands, I suggest you use it: you just need to adapt the IP addresses and get rid of the docker-machine commands.

[Optional] Create a Docker Machine

Do this only if you don’t have a remote machine on hand with which you can play.

So before we can use docker-machine, we need to install it. The detailed instructions can be found on the official page. But basically, you need to download the script from the github repository and chmod +x it.

At the time of this writing, you can find the latest version by doing:

$> curl -L https://github.com/docker/machine/releases/download/v0.12.2/docker-machine-`uname -s`-`uname -m` >/tmp/docker-machine
$> chmod +x /tmp/docker-machine

And then move the docker-machine executable from /tmp to a directory that is in your $PATH.

Basic Working of Docker Machine

Before I explain quickly how we will use docker-machine, let me say that this is a more complex tool that I will present here: I am not writing an article about docker-machine (not yet at least).

Remember that docker-machine is a tool that allows you to provision docker hosts (almost) wherever you want.

What does this last sentence mean? Let’s examine it:

  • “docker-machine is a tool”: so it’s something that you install on top of docker. It doesn’t come with it, and you don’t need it. It’s something that helps you, but it is not mandatory.
  • provision docker hosts”: means that it installs a docker daemon, effectively making a “new” “host” (remember “host” is a “computer” / machine on which runs a docker daemon)
  • “(almost) wherer you want”: that means that you can provision these hosts (install a docker daemon) on your own machine, on remote server, on cloud providers (like Amazon, Azure, DigitalOcean, etc.)

Okay Got It, How Are We Going to Use It?

Here, we’re only going to use it to create docker hosts locally, i.e. on our own computer. Basically, the way to use it is very simple: we use a command to create the virtual hosts. And then when we want to run a docker command in a particular host, we first “point” the docker-machine CLI to the desired virtual host (this is just configuring some ENV variable) and run our command.
It’s as simple as that, really.

So in our case, you can think of docker-machine as a sort of “remote” which you points to the desired virtual docker host before running a command. And boom, that’s it.

So don’t worry: this article will still be very easy to follow, and if you don’t use docker-machine because you have another computer / remote server, the commands are the same ; you just need to ignore the “make the docker-machine point to the desired host”.

Let’s go back to setting up our overlay network!

Set Up a Key-Value Store

The first thing we need to do in order to set up an overlay network is create our KV-store, which will hold the information about the overlay network, which hosts are part of it, which containers in each host, the name and IP addresses of the networks, etc.

Docker basically supports three KV-store:

I’ve only used Consul, so this is the one I am going to use here. Consul (as most KV-stores anyway) is a distributed KV-store, which means it is meant to be used with several instances. All theses instances form a sort of cluster in which the information is stored, with redundancy so that some of the Consul instances in the cluster (some “nodes”) can fail (i.e disconnect, crash, die) without the information being lost.

In our case, we will use docker to install consul. Indeed, rather than installing consul directly on our machine, we will containerize it and run Consul inside a docker container. This way, at the end of this article, you just have to destroy your containers and nothing is messed up in your system, cool no?!

So now we create a virtual docker host with docker-machine, the command is: docker-machine create.

As I said, docker-machine is very powerful and can provision hosts in a lot of different places, so you need to specify where you want to create the host. In docker-machine jargon, this called a driver and you specify it with the -d, --driver option. Here, we’ll use driver virtualbox. This will actually use virtualbox to run a small Linux host system in which docker engine is installed (you need to have virtualbox installed on your system, refer to you distribution for the install (sudo apt install virtualbox for Ubuntu, nix-env -iA nixos.virtualbox for nixOS, etc.)).
We will call this machine kv-store. So open a new terminal (you will undertand why soon) and run:

$> docker-machine create -d virtualbox kv-store
Running pre-create checks...
Creating machine...
(kv-store) Copying /home/nschoe/.docker/machine/cache/boot2docker.iso to /home/nschoe/.docker/machine/machines/kv-store/boot2docker.iso...
(kv-store) Creating VirtualBox VM...
(kv-store) Creating SSH key...
(kv-store) Starting the VM...
(kv-store) Check network to re-create \if needed...
(kv-store) Found a new host-only adapter: "vboxnet1"
(kv-store) Waiting for an IP...
Waiting for machine to be running, this may take a few minutes...
Detecting operating system of created instance...
Waiting for SSH to be available...
Detecting the provisioner...
Provisioning with boot2docker...
Copying certs to the local machine directory...
Copying certs to the remote machine...
Setting Docker configuration on the remote daemon...
Checking connection to Docker...
Docker is up and running!
To see how to connect your Docker Client to the Docker Engine running on this virtual machine, run: docker-machine env kv-store

After this, we have essentially created a new (virtual) host, in which a docker daemon is running. We can list the hosts provisionned with docker-machine with docker-machine ls.

Let’s see:

$> docker-machine ls                                                                            ~
NAME       ACTIVE   DRIVER       STATE     URL                         SWARM   DOCKER        ERRORS
kv-store   -        virtualbox   Running   tcp://           v17.06.1-ce   

So this command lists the provisionned docker hosts, and for each of them, it displays:

  • the name of the host
  • whether it’s active or not (more on this just in a minute)
  • the driver used (remember, it’s to differentiate between a local host created inside a virtual box, or a host provided on a remote cloud provider)
  • the state (whether it’s running, stopped, started, etc.)
  • the URL at which we can join the host (here it’s local, but it can the IP address of a remote server)
  • whether it is part of a Docker Swarm (this is a very interesting topic, but for a later Part!)
  • the version of the docker host
  • whether there are errors

Remember that what we have essentially done here is create a “new server” (which in our case, is inside a virtualbox). But now that we have it, we need to do things with it ; exactly as if we were logged on this machine.
We could use the virtualbox commands to actually log into the machine and then run the commands. But that would be very cumbersome. Instead, you need to remember one of the first things we have understood about docker and the way the docker client works.

When you type docker commands in your terminal, you are using the docker client (docker). The docker client doesn’t do much: it parses your input, check the mandatory arguments, etc. but then it sends all of this to the docker daemon (dockerd) which does the heavy work.

Where does the client send its data?
The docker daemon listens to a socket, and the location of this socket can be configured. In all cases we have seen until now, your docker daemon created a socket at /run/docker.sock, which is a well-known location that the client knows about.

I suppose you have guessed it now, but we are going to use the exact same mechanism to talk to our provisionned docker hosts, and for that we’ll use … guess what … the URL we’ve just seen.

So basically that’s it: you can provision as many hosts as you wish with docker-machine, and then just by setting the correct URL you can have the docker client send the commands to it instead of your local docker daemon. And then you have nothing else to do.
Isn’t that cool?

Now, if you’re astute you must begin to wonder / fear that you will end up chaining docker-machine ls commands, somehow fetching the URL (either an option to docker-machine ls or ugly grepping), somehow chaining the address the docker client will try to contact and then issue your actual docker command.
This looks cumbersome, and it is…

Fortunately, the docker devs have thought about about this and they provided a subcommand env that displays the commands necessary to set up the docker client’s environement. You use it like this: docker-machine env <host-name>.

Let’s see what it gives with our container:

$> docker-machine env kv-store

export DOCKER_HOST="tcp://"
export DOCKER_CERT_PATH="/home/nschoe/.docker/machine/machines/kv-store"
export DOCKER_MACHINE_NAME="kv-store"
# Run this command to configure your shell:
# eval $(docker-machine env kv-store)

As you see, it sets some variables, the most important being DOCKER_HOST, and you see that it matches the URL field we saw previously. You surely noticed the export clauses, and at the end, there is a comment that explains how to use: you need to use eval which is a shell’s command, used to set / change environment variables.

So when you have several provisionned hosts, you will switch between them with these eval $(docker-machine env <host-name>) commands, so it’s a very simplified procedure.
Always make sure you know which (if any) host is active before issuing docker commands, you can check which host is active with docker-machine active which returns the name of the active host.

$> eval $(docker-machine env kv-store)
$> docker-machine ls
NAME       ACTIVE   DRIVER       STATE     URL                         SWARM   DOCKER        ERRORS
kv-store   *        virtualbox   Running   tcp://           v17.06.1-ce   

As you see, the ACTIVE field now contains a star which indicates that this host is the one that will receive the docker commands now.

Now that we have made a host active, you can see the effect by running traditionnal commands:

$> docker images
REPOSITORY          TAG                 IMAGE ID            CREATED             SIZE

$> docker ps -a
CONTAINER ID        IMAGE               COMMAND             CREATED             STATUS              PORTS               NAMES

As you can see: there is nothing. And this is normal: the host has just been created, so no images yet, no containers, nothing.

“Hey, but now I can’t run ‘normal’ docker commands anymore, I can’t see the containers on my machine!”

This is normal: your environement is configured so that your docker client sends the commands to a host provisionned by docker-machine. So you just need to unset this, so that it falls back to your local docker daemon.

In order to do this, you need to use the -u, --unset option of docker-machine env. See:

$> eval "$(docker-machine env --unset)"

We can tripple check that our commands worked and we are “back to normal”:

First, check with docker-machine active:

$> docker-machine active
No active host found

Then, check with docker-machine ls

$> docker-machine ls
NAME       ACTIVE   DRIVER       STATE     URL                         SWARM   DOCKER        ERRORS
kv-store   -        virtualbox   Running   tcp://           v17.06.1-ce   

(we can see the star * disappeared and was replaced by a dash -).

And finally, we can test to run our docker commands:

$> docker images
REPOSITORY          TAG                 IMAGE ID            CREATED             SIZE
nginx               latest              db079554b4d2        6 months ago        182MB
ubuntu              16.04               104bec311bcd        8 months ago        129MB

Now you know the basics of how to interact with docker-machine and the provisionned hosts. Make sure you got it, as we will use it from now on.

So! Back to our topic, we were about to set up a KV-store, so let’s do that, shall we?

First, we need to activate the kv-store host with:

$> eval "$(docker-machine env kv-store)"
$> docker-machine active

Now we run the progrium/consul docker image, to create a consul server:

$> docker run -d -p "8500:8500" -h "consul" progrium/consul -server -bootstrap

It will tell you that it cannot find the image locally, then it will download it and run it. You should have an output similar to:

$> docker run -d -p "8500:8500" -h "consul" progrium/consul -server -bootstrap                    ~
Unable to find image 'progrium/consul:latest' locally
latest: Pulling from progrium/consul
c862d82a67a2: Pull complete
0e7f3c08384e: Pull complete
0e221e32327a: Pull complete
09a952464e47: Pull complete
60a1b927414d: Pull complete
4c9f46b5ccce: Pull complete
417d86672aa4: Pull complete
b0d47ad24447: Pull complete
fd5300bd53f0: Pull complete
a3ed95caeb02: Pull complete
d023b445076e: Pull complete
ba8851f89e33: Pull complete
5d1cefca2a28: Pull complete
Digest: sha256:8cc8023462905929df9a79ff67ee435a36848ce7a10f18d6d0faba9306b97274
Status: Downloaded newer image for progrium/consul:latest

Let’s just check the consul server is actually running in our kv-store docker host:

$> docker ps
CONTAINER ID        IMAGE               COMMAND                  CREATED             STATUS              PORTS                                                                            NAMES
5af66f2f99d6        progrium/consul     "/bin/start -serve..."   53 seconds ago      Up 51 seconds       53/tcp, 53/udp, 8300-8302/tcp, 8400/tcp, 8301-8302/udp,>8500/tcp   frosty_northcutt

Yes, so far so good.

Configure the Docker Daemon Hosts

Before we move on, I’d like to make a small recap of what is going on, because that is the end of the article (which is growing in size), I have just made quite a detour to introduce basic docker-machine commands and I feel I’m losing you…

So here we go, small recap!

In the first (big) part of this article, we learned about Docker Networks and how it can be used to make containers communicate together. This allowed us to do great things and really unlocked a new aspect of Docker that we were previously missing.
All of this is very good, but it is somewhat limited by the fact that you can only create Docker Networks on your machine and connect local containers to it (understand: containers that are running on the same machine), and we found that if would be so great to extend this exact same mechanism (so we don’t have too much new to learn) so that it worked between machines that are remotely connected.

This is possible through the use of Overlay Docker Networks.
But we understand that if we want to have several remote machines talk to each other and maintain a common state, we need a shared medium to synchronize data. For instance, which containers and on what machines are in which Overay Docker Networks, what is this container’s internal IP, etc.

This is achieved by Docker by leveraging the power of distributed Key-Value stores, such as consul, the one we use here.

In the last section (Set Up a Key-Value Store), we did two things:

  1. we learned how to use docker-machine to create virtual hosts on our own machine, to emulate the fact that we have several (physical) servers / hosts (people who actually_ have several servers / hosts can bypass this part obviously)
  2. we actually created one instance (server) of the consul KV-store.

So that’s it for the recap: we now have one instance of consul running, and it’s reachable at the IP address given by docker-machine ls. Oh and by the way, if you need to have just the IP address of a host provisionned by docker-machine, instead of running docker-machine ls and grepping, you can use docker-machine ip <host-name> and it returns just the IP address, which is pretty handy ;)

A last note: consul (and KV-stores in general) are much more complex systems that we made it appear here. This is completely different and independant from docker: you can use KV-stores to synchronize data across several machines. There is a whole topic about distributed KV-store.
In our case, we run just one instance of the KV-store (in the host named kv-store provisionned by docker-machine), but this is dangerous in real-life applications: if this machines fails and the KV-store cannot be reached anymore, you lost your information.
This is why distributed KV-stores are generally run in several instances, on several different machines and these machines communicate together to maintain a distributed set of information, so that if a machine (or a number of machines) fail, the information can still be obtained.

We don’t do that here because the goal of the topic is to teach about (Overlay) Docker Networks, but if you are going to use Overlay Docker Networks in your production applications, I strongly suggest you document about consul (or whatever the KV-store you chose) and learn to properly set up a really distributed cluster, so that you don’t lose your information the day one of your machines crashes.

Okay! So what’s next?!

Well now that we have a KV-store up and running, we need to configure the docker daemon(s) that should be using it! Up to now we did not have anything to configure because when we created Docker Networks and connected containers to them, this information was kept locally by Docker and it did not need to be synchronized.

So now we must add two options to our docker daemon:

  1. one IP address where it can reach a KV-store (and what type of KV-store it is)
  2. one IP address / network interface on which it should advertise. What does that mean? It’s easy: suppose we have several docker daemons (on several machines) that should communicate together. Meaning we will have Overlay Docker Networks created, some containers on these machines and some on these containers will be part of the Overlay Docker Networks. But the docker daemons need to talk to each other to synchronize runtime data (i.e. “can you resolve this particular IP address, because it doesn’t belong to any of my containers”, etc.) Well they do so through the “advertised” IP address. Easy!

In order to configure the docker daemons, we have several ways, I will talk about two:

  1. one quick but dirty way (most useful to test if the configuration is okay)
  2. the modern, clean way

As you may know, the docker daemon is now run with dockerd, which accepts command-line parameters. The two parameters we want to set now are (in order):

  1. cluster-store
  2. cluster-advertise (be careful, this is “advertise” and not “advertize”).

The quick way is simply to launch dockerd by hand with the options, the clean way is to create a systemd unit file, and set the configuration in it.

When you want to configure your own docker daemon, you first need to stop it (be warned that it will stop all of your containers). If you had launched it with dockerd previously, then hit CTRL + C, if you had launched it with systemd, then type systemd stop docker.

But in our case we are not going to set up our local docker daemon (we could, but if you forget to remove the setup later, it will mess up your daemon, and you might have a tough time remembering it), instead we will provision hosts with docker-machine and configure their docker daemons.

Now we are going to set the parameters.

We need to set the IP address of our consul server, which is the IP address of the provisionned docker host kv-store so let’s get it:

$> docker-machine ip kv-store

(This is obviously subject to change on your machine, use your own returned IP address).
Now we need to set port on which the consul server listens ; we configured it to be 8500 in the previous docker run command (scroll back up, you will see -p 8500:8500).
And lastly, we need to tell docker that it’s a consul cluster, and not a zookeeper or etcd. So the full parameter for the first option is: --cluster-store=consul://

That was not too hard, right?

The second option is even simpler, it takes the form: <ip-or-interface>:<port>. Remember that this is the IP:port that other (potentially remote) docker hosts will contact your host at. So make sure this IP:port combinaison is open and reachable from your remote servers.
The port must be 2376 (this is docker’s protocol).

So you will have something like --cluster-advertise=eth1:2376 in our case (the hosts provisionned by docker-machine has an interface called eth1), adapt if you need.

Creating And Using Overlay Docker Networks

Now that we have our options, if you were to configure your local docker daemon you could either launch the daemon by hand, like this: dockerd --cluster-store=consul:// --cluster-advertise=<interface>:2376 or edit the systemd configuration file to add the options on the ExecStart= line. When it is started, nothing will look different than before, but now the daemon would know where to join a KV-store and thus the Overlay Docker Network feature would be enabled.

Let’s create our first Overlay Docker Network.

As I said earlier, everything we saw with traditionnal Docker Networks still applies ; Overlay Newtorks are built upon this. So we still create an Overlay Docker Network with docker network create, but now we add the -d, --driver option to tell docker that is not “bridge” anymore (it’s the default we have been using until now) but “overlay”.

Just before we do that, though, we will provision another host with docker-machine, whose docker daemon we will configure (in order not to touch ours).

Provisionning Another Host

For now we have all in all two hosts:

  • our local machine, the one on which we have been experimenting with until now and which we won’t modify
  • the kv-store (virtual) machine that we provisionned with docker-machine

Okay, that is all nice and all, but:

  • kv-store doesn’t really count because we only created it in order to have an instance of consul running. We could totally run some other containers on it, but I’m afraid this might confuse some of you. So for the sake of our little test setup, let’s forget about kv-store.
  • it leaves our own docker daemon, on our local machine. We could also run containers on our local docker daemon to take part of the overlay docker network, but as I have explained a few screens up, in order for a docker daemon to work with overlay networks, you need to configure it. So we would need to configure our docker daemon. And then there’s a risk you forget about the configuration and later find that your docker daemon has troubles starting or is very slow because of this. So we won’t use our local docker daemon for the sake of this test.

This leaves us with basically 0 hosts to make our test. So we will create two of them with docker-machine and then problem fixed! Easy!
Note: you can totally use your own docker daemon, it works the same, make sure to remove the configuration when you’re done!.

Let’s do this!

As this will be the first host we will actually make tests on, we will naturally call it first-host (I have a talent for naming…). Before I give the command, I want to clarify one or two things.

So we will provision a virtual host with docker-machine. If the Docker dev team went to all the trouble of creating this docker-machine tool, you can imagine they tailor-made it to suit docker needs.
So when you provision a new host with docker-machine and the -d virtualbox driver, it’s not just a simple Linux host ; the whole goal of docker-machine is to install and configure the docker daemon automatically.

Do you remember the configuration we have talked about regarding how we would set up our own docker daemon (namely setting --cluster-store and --cluster-advertise)? Well, we will obviously need to do this for the provisionned hosts too, nothing is magic: the automatically installed docker daemon still needs to know about the KV-store.

But fortunately, we won’t need to log into the virtual host and edit the configuration by hand: docker-machine has an option, --engine-opt which is used to pass configuration options to the docker daemon it will automatically install.

So now we can pass these two options. We need two things:

  1. the IP address of the KV-store, but that is easy, it’s the IP address of kv-store, the provisionned host. Remember that we can get it with docker-machine ip kv-store, so we will use this directly in the command
  2. the IP / interface on which to advertise the docker daemon. I will make it easy for you: when you provision a docker host with docker-machine, then the virtual host has a (virtual) network interface eth1. So you can use this.

Enough talk, let’s do it now!

$> docker-machine create -d virtualbox --engine-opt="cluster-store=consul://$(docker-machine ip kv-store):8500" --engine-opt="cluster-advertise=eth1:2376" first-host

Running pre-create checks...
(second-host) Default Boot2Docker ISO is out-of-date, downloading the latest release...
(second-host) Latest release for github.com/boot2docker/boot2docker is v17.07.0-ce
(second-host) Downloading /home/nschoe/.docker/machine/cache/boot2docker.iso from https://github.com/boot2docker/boot2docker/releases/download/v17.07.0-ce/boot2docker.iso...
(second-host) 0%....10%....20%....30%....40%....50%....60%....70%....80%....90%....100%
Creating machine...
(second-host) Copying /home/nschoe/.docker/machine/cache/boot2docker.iso to /home/nschoe/.docker/machine/machines/second-host/boot2docker.iso...
(second-host) Creating VirtualBox VM...
(second-host) Creating SSH key...
(second-host) Starting the VM...
(second-host) Check network to re-create \if needed...
(second-host) Waiting for an IP...
Waiting for machine to be running, this may take a few minutes...
Detecting operating system of created instance...second
Waiting for SSH to be available...
Detecting the provisioner...
Provisioning with boot2docker...
Copying certs to the local machine directory...
Copying certs to the remote machine...
Setting Docker configuration on the remote daemon...
Checking connection to Docker...
Docker is up and running!
To see how to connect your Docker Client to the Docker Engine running on this virtual machine, run: docker-machine env first-host

We can check that we have a new host:

$> docker-machine ls

NAME          ACTIVE   DRIVER       STATE     URL                         SWARM   DOCKER        ERRORS
kv-store      -        virtualbox   Running   tcp://           v17.06.1-ce   
first-host    -        virtualbox   Running   tcp://           v17.07.0-ce   

(Funnily enough, it seems to have updated the docker version between the time I created first-host!).

Time to Create the Overlay Network

So now we can execute the command we have talked about, let’s first activate first-host: eval $(docker-machine env first-host) and run the command:

$> docker network create -d overlay TestOverlay

And then you can check with

$> docker network ls
NETWORK ID          NAME                DRIVER              SCOPE
2c5c37a080f5        TestOverlay         overlay             global
2d610b04d998        bridge              bridge              local
2683c3824f8f        host                host                local
a84e731c4387        none                null                local

Now you should spot the new network which differ because its DRIVER option now lists overlay and its SCOPE is global.

So far so good, but nothing too fancy. The magic will come when we add a second host. Let’s create second-host right now with the same command: docker-machine create -d virtualbox --engine-opt="cluster-store=consul://$(docker-machine ip kv-store):8500" --engine-opt="cluster-advertise=eth1:2376" second-host

Now I am going to print a startled cat image because I want your attention:

Okay, now I’ve got your attention, please follow what we are about to do, and what happens.

First, we are going to activate the second-host host with eval "$(docker-machine env second-host)".

Let’s make sure it worked:

$> docker-machine active

Okay, so now our docker client “points to” (i.e. send its commands to) the docker daemon inside second-host, which let me remind you must be considered like a total other machine, one that might be on a server provider 10,000 km from your.

As we would expect from this newly-provisionned docker host, there is nothing (no containers, no images, no volumes, nothing):

$> docker images
REPOSITORY          TAG                 IMAGE ID            CREATED             SIZE

$> docker ps -a
CONTAINER ID        IMAGE               COMMAND             CREATED             STATUS              PORTS               NAMES

$> docker volume ls

Let’s check the networks for completeness:

$> docker network ls

NETWORK ID          NAME                DRIVER              SCOPE
2c5c37a080f5        TestOverlay         overlay             global
2e745f1f19da        bridge              bridge              local
6cf2d0234769        host                host                local
bbec139d7ee0        none                null                local

So we’ve got our traditional 3 basic networks (none, host and bridge) and that’s about … wait what?!

(hint: you should be very excited now).

So yes, what just happened is very cool: we have essentially created a new host with a fresh-new docker installation, but somehow it already knows about the TestOverlay Docker Network that we created:

  1. earlier
  2. on a completely different machine!

Obviously, the somehow is explained easily: the docker daemon contacted the KV-store thanks to the --cluster-store option we passed earlier.

But the very cool thing to take away from this, is that by configuring a KV-store and setting up the docker daemons correctly, we can have Docker (Overlay) Networks shared among completely independant machines (hosts)!

Please, please take a few seconds (minutes if you need) to re-read this last sentence and make absolutely sure you understand the consequences and the power it will give us.

Seriously, I know it seems like I do a big deal out of this, but this is a big deal. Think about everything we learned in this Part: breaking the isolation between containers in a controlled manner, embedded DNS server that automatically resolves the containers’ names to their correct IP etc… well everything is still working, and on top of that… it is now spread across several servers.

We have just (almost freely) extended our nice little comfy zone of isolated-yet-communicating containers between servers. Now, from a docker point of view, when you have containers C1 and C2 inside Docker Network N, it doesn’t matter if C1 and C2 are on the same server S1 or if one is on server S1 and one of server S2.

That is awesome and I’m not continuing the article until I hear everyone of you say that it is awesome.

I’m waiting.

Good. Now we can wrap things up.

Creating Containers And Be Amazed

We are not going to make extended tests because as I said several times, the Overlay Networks work exactly the same as Bridge Network. And we have covered Bridge Networks quite extensively.
We will still make one or two basic tests to convince ourselves that it does appear to work, but aside from this, you can apply what you already know.

But still, it deserves a small test nonetheless. We are going to create a container on first-host and another one on second-host, obviously we are going to connect them to the TestOverlay Docker Network and we are going to see if they can talk to each other.

Make sure first-host is active with docker-machine active. If not, make it active with eval $(docker-machine env first-host).

Then create a container, as usual, connecting it to the network:

$> docker run -itd --name cont_one --net TestOverlay nginx
Unable to find image 'nginx:latest' locally
latest: Pulling from library/nginx
94ed0c431eb5: Pull complete
9406c100a1c3: Pull complete
aa74daafd50c: Pull complete
Digest: sha256:788fa27763db6d69ad3444e8ba72f947df9e7e163bad7c1f5614f8fd27a311c3
Status: Downloaded newer image for nginx:latest

You might have remarked that this time I used image nginx and not net-test. The reason for this is that net-test is an image we have created (we have written a (small) Dockerfile and built on our local machine. Since we did not host our Dockerfile on the Docker Hub or in any private docker registry, we cannot easily have this image on the hosts. So we use a standard image nginx which happens to have ping installed natively).

So we created a container named cont_one on the host first-host and it is part of the TestOverlay network. We can quickly check it is indeed part of the network:

$> docker network inspect TestOverlay
"Containers": {
            "398de04064c10b6ef2be606474620aa77c50910497213f723c870eb1a0508f6c": {
                "Name": "cont_one",
                "EndpointID": "a68620105c26a6bf9eb91f7c968fb4987ce7d3cca36c13edd1cc5c7fba63476e",
                "MacAddress": "02:42:0a:00:00:03",
                "IPv4Address": "",
                "IPv6Address": ""

So that’s good: the container is connected.

Now, let’s activate second-host with eval $(docker-machine env second-host).

A quick docker ps shows you that the cont_one is not listed anymore.

What would happen if we ran docker network inspect TestOverlay on this host? Remember this is second-host now, it is not the host on which the container cont_one was created and exists?

Easiest way to find out out is to try:

$> docker network inspect TestOverlay

"Containers": {
            "ep-a68620105c26a6bf9eb91f7c968fb4987ce7d3cca36c13edd1cc5c7fba63476e": {
                "Name": "cont_one",
                "EndpointID": "a68620105c26a6bf9eb91f7c968fb4987ce7d3cca36c13edd1cc5c7fba63476e",
                "MacAddress": "02:42:0a:00:00:03",
                "IPv4Address": "",
                "IPv6Address": ""

So… we do see that it lists a container cont_one that is connected to this network, but if you look closely, the ID is not the same. It starts with ep- which it typical of containers listed that are not present on the current docker host (it’s not important for the moment, just keep this in a corner of your mind).

So this is still pretty interesting: we are on a docker host where container cont_one does not exist, it was never there but still it is listed when inspecting the Docker Network. All of this is possible, you guessed it, thanks to the KV-store that stores information.
You would note however, that the EndpointID is the same on both output, and this is what is used internally by docker to actually identify the container.

Oh and one last thing you might have noticed: the IP address if very different than the IP addresses we have dealt with until now. They used to be in 172.X.Y.Z, but now they are in 10.X.Y.Z. So that is another way of telling you are in an Overlay network.

Now let’s add a container to this network, on this host (second-host)!

$> docker run -itd --name cont_two --net TestOverlay nginx

A quick inspection of the network shows that it worked as expected:

$> docker network inspect TestOverlay

"Containers": {
            "c6a71d242ae2fcdbe4c546af01b84a144c94a28eca1ff2f81dfc3de3927f7a3d": {
                "Name": "cont_two",
                "EndpointID": "3211821c4e5cc07a5c64ef2569d207b5997143b0592df29de4cedd9c7c104d3e",
                "MacAddress": "02:42:0a:00:00:05",
                "IPv4Address": "",
                "IPv6Address": ""
            "ep-a68620105c26a6bf9eb91f7c968fb4987ce7d3cca36c13edd1cc5c7fba63476e": {
                "Name": "cont_one",
                "EndpointID": "a68620105c26a6bf9eb91f7c968fb4987ce7d3cca36c13edd1cc5c7fba63476e",
                "MacAddress": "02:42:0a:00:00:03",
                "IPv4Address": "",
                "IPv6Address": ""

And now’s the real test: we will enter container cont_two and try to ping cont_one, which is on a different host, remember.

$> docker exec -it cont_two bash
root@c6a71d242ae2:/# ping -c 4 cont_one
PING cont_one ( 56 data bytes
64 bytes from icmp_seq=0 ttl=64 time=0.402 ms
64 bytes from icmp_seq=1 ttl=64 time=0.359 ms
64 bytes from icmp_seq=2 ttl=64 time=0.411 ms
64 bytes from icmp_seq=3 ttl=64 time=0.417 ms
--- cont_one ping statistics ---
4 packets transmitted, 4 packets received, 0% packet loss
round-trip min/avg/max/stddev = 0.359/0.397/0.417/0.023 ms

And it works!

So basically, containers cont_one and cont_two are talking to each other as if they were in the same private network (LAN) even though they are on two different hosts (computers) that can be anywhere in the world.

This is fantastic. Seriously, you don’t understand how fantastic this is.

What that means is that you can keep thinking and designing your Docker stacks correctly, i.e. separate the different logical units (processes) in different containers (one container for the web server, one container for the database, one container for the redis cache, one container for service monitoring, etc.), but now you don’t have to cram all of these containers into a single machine…
You can have as many machines (hosts) as you want, you can fine-tune the configuration of each machine to fit its use-case and on top of this, “all you have to do” is configure your docker daemons and your KV-store correctly, and thanks to this, you can create Docker Networks across these machines and connect the containers to them.
The containers are just contacting other containers by using their container names, the embedded DNS server takes care of the routing for you and you can essentially keep thinking and designing your stack as a big LAN network (minus the network latency of course).

This is incredibly useful.

I mean a small overview of this might be something like: you buy a server which is optimized for bandwidth but doesn’t have lots of disk space to put all your nginx containers on it, you can buy a server which has a lot of RAM to put your redis cache on it and you can buy servers with a lot of disk space to put your postgresql containers on it.
Once you have done that, configure the docker daemons, create your Overlay Docker Networks, connect your containers and This. Just. Works.
Everything is abstracted for you.

You don’t have to take care of routing, of matching container name to the IP address of the host nor macthing the container’s IP address inside a host.
All of this is done for you, by docker. This is great, really.

Gotchas And Common Errors

As I said before, Overlay Docker Networks work exactly as Bridge-based Docker Networks as far as the user (it’s you is concerned), provided you managed to setup the system. I want to give you a few hints to diagnose your setup in case it doesn’t work.
It’s not much, but I found that it was very easy to avoid these and a good part of the time, it solved the problem.

So here they are:

  • The first obvious thing to know / remember is to have set up your docker daemons to work with the KV-store, for that, you need to have the two options --cluster-store with:
    1. a protocol (for instance consul://)
    2. an IP address (the IP address of one of the KV-store node, make sure this address is reachable by your host)
    3. the KV-store port (for consul this is 8500 if you did not change the default configuration) Make sure the IP address and the port are opened by your firewall. The second option is --cluster-advertise which needs the network interface (eth0, wlan0, etc.) that your docker host will use to advertise its presence, and the port (this is 2376 if you did not change Docker’s settings)
  • If you set these up with a systemd unit file (most likely /usr/lib/systemd/docker.conf.d/docker.conf), then make sure you added a newline at the end of your file. I have spent countless hours reviewing everything: the addresses were correct, the ports were correct, the access were authorized, everything should be okay yet nothing worked, etc. And then I started “looking the keys inside the fridge” and one day I opened the systemd file with another editor which automatically added the newline and it worked. So… always, always end your config files with a newline at the end. This one, I guarantee will save you hours!
  • Make sure to look at the docker daemon logs (either by starting it by hand, or looking at journald), but the logs might indicate that something’s off. For instance, if you get the --cluster-advertise option wrong, you will see a message like this:
WARN[0001] Multi-Host overlay networking requires cluster-advertise(x.x.x.x) to be configured with a local ip-address that is reachable within the cluster

And that indicates that it doesn’t work.
- Another thing that is worth knowing is that the two options --cluster-store and --cluster-advertise are both necessary for Overlay networks to work, but they are not related to the same thing, and they work independently of each other.
To make it short, --cluster-store allows the docker daemon to contact the KV-store and fetch information. This is what makes docker network ls shows the Overlay Docker Networks (if you misconfigured it, only the Bridge-based Docker Networks will show up), and this is what makes docker network inspect <name-of-overlay-network> shows the information about which containers are connected, their IP addresses inside the network etc. So even if you only configure --cluster-store or if you misconfigured --cluster-advertise, this will still show you and you might think that everything is working as it should.

The second option, --cluster-advertise is what makes the docker hosts able to communicate and synchronize together. This is what makes containers in different hosts able to talke to each other. It handles the routing. So if you have configured --cluster-store together but --cluster-advertise is misconfigured, then when trying to ping a container in another host, you will get 92 bytes from 523ed8b20522 ( Destination Host Unreachable for instance.
This is very important because you need to be able to diagnose your problem.

The TL;DR is that:

  • if you don’t see the Overlay Networks with docker network ls, then it’s a problem of --cluster-store (so either it points to the wrong IP address, the wrong port, or the IP address / port are not opened by the firewall, or the KV-store is simply not running, etc.)
  • if you can see the Overlay Networks but your containers fail to talk to each other (Destination Host Unreachable, No route to Host, Timeout), then it’s a problem of --cluster-advertise (wrong network interface, wrong port (2376), the network interface is not opened on the outside, firewall blocking (port 2376), etc.).

Security Concerns

A last but Very important part remains: security. I love security and I find this is a fascinating topic. However this article is too long already to dwell too long on this. But I have to say a word or two about this.
In case you are using only Bridge-based networks like we did in the first part of this article, you don’t need to worry too much about security: since everything is on your computer (containers are all on teh same host, they communicate via iptables on the host), it’s fine. Nothing is leaked.
Just make sure that if you open ports, you don’t open them on the outside by accident (remember that -p 8080:8080 will bind the container’s 8080 ports to the host’s 8080 ports on all interfaces whereas -p do the same, but only on the host's10.20.3.4` network interface).

If you are using Overlay Docker Networks, however, the story is a bit different. As we have seen, we need instance(s) of a KV-store (consul, ZooKeeper, Etcd, etc.). Even though we only used one server of consul here, I told you that in real setups, you should definitely have several, so that you have redundancy. These instances must communicate together over the Internet. And by default, the communication if not encrypted. So it’s completely unacceptable.
In the same spirit, your docker daemons must communicate with the KV-stores (that’s the --cluster-store option) and together (that’s the --cluster-advertise option). And these communications are also unencrypted. This is dangerous and unacceptable for anything else that basic tests (like in this article).

You have several possibilities. As far as KV-stores are concerned, if you read their documentation, you should find instructions about how to setup a cluster of consul over SSL / TLS (it’s also obviously true for the other KV-stores).

As for communications between docker hosts, it depends on your setup. In a real, solid and serious setup where you rent servers on a server provider only, you can usually find servers that are part of an internal private network. Meaning that the servers have two network interfaces, let’s say eth0 and eth1. eth0 is the public network interface, the one connected on the Internet, and eth1 is not connected to the Internet, but is part of an Internal, fast internal network. And you can use this private eth1 interface to connect to your other server, securely. This way you don’t need to encrypt the communication between the docker daemons, because it’s not reachable from the outside.

But this only works between servers on the same provider and that are part of this network. When it is not the case, I found that a relatively good, simple and easy-to-implement solution was to create a VPN tunnel between the hosts with the docker daemons. So basically on your docker daemons you would have tun0 network interfaces, which you would use to --cluster-advertise. Obviously you would have set up your VPN tunnel with encryption, so this way the communication if encrypted.

You can also do the same for the communication between the docker hosts and the KV-stores.

What is also possible for these things is to use nginx as a (reverse) proxy. On each each, nginx can communicate with the servers (either the docker daemon or the KV-store) on an internal network interface (localhost) and proxy the answers over a TLS/SSL encrypted communication.

Anyway, this is a topic in itself so I can’t cover it in just a paragraph, but make sure you take this into accounts ; because otherwise you will have problems!

Here you go, with these sets of indications, I’m confident that you should be able to diagnose the vast majority of the problems you might be having with Overlay Networks, and if you still can’t, either , and go to #docker and I’m here (@nschoe).

And once you managed to setup the Overlay Networks, with them and everything we saw in the first 4 Parts, you should be able to make very nice container stacks.


Ok-kay! So here we are: we have finally reached the end of this Part V about Docker Networks. I know it was long, and I know it took a lot of time to write (and thus publish), and I’m sorry about that.
But Docker Networks are the last fundamental Docker concept. Now you truly have everything you need to build very complex stacks The Right Way.

Also, I found (hanging on IRC and reading Stack Overflow questions) that it was (as with the other fundamental Docker concepts) greatly misunderstood and misused. This is why I wanted to make an article that explained the Docker Networks the best I could and hopefully clear any misconceptions about them. Even if it meant taking my time.

So what did we learn in this Part V?

We learned to break the isolation between containers that we have been talking about since Part I, but in a controlled way. In other words, we learned how to open a small door between chosen containers so that they can know about each other, and then communicate, exchange data, etc.
All the other aspects are still completely hidden (a container still can’t know about another container’s processes, resources, filesystem, etc.), and obviously only the containers we chose can know about each other. So we have the best of both worlds.

Part V is also a sort of milestone, because with it we have finsihed learning about the fundamental concepts of Docker, we know all the “basics”. By “basics” here I don’t mean “basic usage”, as in “simple”, but rather the core concepts, the things that we need to make real use of Docker. Let’s recap them briefly:

  • In Parts I and II we learned about Docker itself, what this is (because it’s always important to know what we are talking about) and we learned about the difference between an image and a container: an image is a “recipe” to setup an environment and a container is one instance of it. We also learned that containers should be ephemeral, they are designed to be short-lived: A. Container. Is. Just. A. Process.
    We always should be able to stop, destroy and rebuild a container.
  • In Part III, we learned about how to write Dockerfiles and how to build our custom images, with this we can now design a tailor-made environment that suits the software you want to run.
  • Until Part IV we had a problem: if we were making short-lived containers, containers that could be shut down at any moment, even destroyed, we would lose data. So we needed to have a way to “save” data somewhere. This is where Part IV comes with Docker Volumes.
    We learned how to make data persistent by storing it in Docker Volumes (which have a different life cycle than containers), and we also lerned how to share data between the host and the containers.
    I emphasized on the fact that even though storing data in a volume and sharing data with the hosts were both grouped under the name “Docker Volumes”, it was two different things and one should not be used in the place of the other.
  • But even then we still had a problem: we were able to make ephemeral containers (“just processes”, remember?), we could store our important data in Volumes, but if we wanted to have only one (logical) process per container, we were limited to very simple setups, because if our setup involved several containers, it was impossible to make them communicate one with another. Part V solves this by introducting the concept of Docker Networks, which -as the name implies- is a way to make several containers communicate with each other.
    We then extended this concept to several machines, so that distant, remote servers running a docker daemon (and running containers of their own) could easily have networked container, basically for free (“for free” here means that we don’t have anything new to learn and almost nothing new to setup).

Now, after those 5 Parts you can do (almost) anything, and you can sure make very complex stuff. For information, I have been working with Docker for a couple of years now, we have a pretty complex client stack and they rely on what we have been learning on these first 5 Parts.

So feel free to experiment now, build complex stacks, and enjoy.

So This is It? That’s How It Ends?…

I mean… for this article, yes, but not for the series, no! Fear not.
We are not done with Docker yet, not even close, we have lots of cool stuff to discover, and I want to keep writing about them (as long as you still want to read!).

There are in particular two very important topics that I want to discuss, and I’m sure you have heard of them, they are:

  • Docker Swarm, and
  • docker-compose

Both of them are very powerful tools, which open your world of possibility even more, but they are (alas) too often, much too often confused with what I call “vanilla docker”. They sound attractive (and they are!) but beginners almost always jump to these tools too early. And then they must deal with all the notions we have covered in these 5 Parts, plus docker-compose and Docker Swarm.
As long as they follow “click-here, do this” tutorials, it’s fine, but when they start having problems on their own, they can’t diagnose where the problems comes from.

This is why I haven’t talked about any of these tools yet. I tried to give you all the tools that I could think of to really understand what docker does, what docker is composed of and what docker can do. Now that you have these tools, I feel that we are ready to talk about Docker Swarm and docker-compose. So I think the next articles are going to be about them.
I’m not sure of the order yet, so I prefer not to give that away (in fear of disappointment), but “It Is Coming.”.

In the mean time, I hope you will play with Docker and what we have learned together in these first 5 Parts, and I urge you to email me if anything is still not clicking, or if you feel the way I introduced a topic is still misleading (also I take “thank-you” email as well!).
If you have any topic that you would like me to cover in upcoming Parts, also feel free to suggest by email.
Also, last piece of news: I’m thinking about going freelance and take on some consulting missions regarding docker (maybe advise people on the way to build their stacks, maybe actually build the stack, review some already existing stack, etc.).

If you think you might be interested, don’t hesitate to contact me. I still don’t have the official status, etc, it’s something I have been thinking about for quite some time. Now that this series is taking some traction, I feel it would be a good time. Anyway, for whatever reason, you can contact me on .

I hope you liked the article, and see you in Part VI!

September 12, 2017