Guest Blog By Jérôme Petazzoni | Docker Tinkerer Extraordinaire
Have you heard about dotCloud? If you haven’t, I’m going to give you a hint: it is a PAAS company. Another hint: eventually, dotCloud open-sourced their container engine. That container engine became Docker.
This is a quasi-archeological account of some of the early design decisions of dotCloud, some of which have shaped how Docker is today (and how it is not). “How is this relevant to my interests?” you ask. If you are not using containers, and not planning to, ever, then this article will not be very useful to you. Otherwise, I hope that you can learn a lot from our past successes and failures. At the very least, you will understand why Docker was built this way.
First of all, a disclaimer
Don’t consider this as a set of guidelines, recommendations, or whatever. It’s important to keep in mind that when dotCloud was created (and for quite a while!), things were very different:
- EC2 had just rolled out support for custom kernels
- EBS had major outages at least once a year
- Linux 3.0 wasn’t out yet
- Consul and etcd didn’t exist
- Go 1.0 hadn’t been released yet
This might help to get some perspective on some of our technical choices.
Take everything I say here with a grain of salt. I no longer have access to the original dotCloud code, and while I knew that codebase pretty well, I don’t have an eidetic memory and it’s very possible (and even likely) that I misremember a few things. If you were there, and think that I got something wrong, let me know! I’ll be happy to fix it.
And then, a short page of boring history
At $STARTUP_NAME , we always knew that containers were the future, and we were using them before they were cool! We are true containhipsters and we are glad that everybody else is finally seeing the blinding light that we saw decades ago!
This is not a cheap shot at Bryan Cantrill, for whom I have an inordinate amount of respect and admiration. Sometimes I wish I had been born earlier (and also smarter), and got a chance to work on Solaris. Now that’s a decent OS (even if the userland is extremely picky about who it makes friends with), and when the Joyent crew runs containers, they’re not messing around. (If you want a taste of what it’s like to run containers in production like a boss, check out this talk, you’ll see what I’m talking about.)
But no Solaris for me! Instead, a friend whose hair and beard could rival with Stallman’s gave me a Slackware CD in the mid-90s, and I’ve been stuck with Linuxever since. (I tried FreeBSD once. I managed to crash the installer and then went onto file one of the most inane bug reports ever.)
Fast forward to 2008, when fellow hacker Solomon Hykes gives me (and others) a demo of dotCloud. Back then, dotCloud was a CLI tool allowing to author container images, move them around, and easily instantiate them on multiple machines. That demo was honestly pretty similar to the one in Solomon’s lightning talk at PyCon in 2013, but the tech behind it was very different. And for a good reason: what Solomon demo’ed in 2013 was the result of 5+ years of trial, error, and learning hard truths the hard way.
This is what our first containers looked like!
The dotCloud container engine (the ancestor of Docker) started as a Python CLI tool called dc . (Yes, we knew that it conflicted with the old-school desk calculator program. No, we didn’t care.)
dc acted as a frontend to LXC and AUFS. Specifically, dc could:
- manage container images (pull/push them from/to a registry),
- create a container using one of these images (leveraging AUFS copy-on-write),
- configure the container, by allowing any file to be generated from a template (for instance, putting the correct IP address in /etc/network/interfaces),
- start the container, by automatically creating its LXC configuration file and invoking lxc-start ,
- dynamically expose ports, by managing a set of iptables rules,
- and a few other cool things.
This is what interacting with dc looked like. Keep in mind that I haven’t used dc in 3 years and I don’t have the code anymore, so this is only approximate.
So far, so good.
There are a few profound differences, however, between dc and modern Docker.
We wanted to be able to track and audit accurately changes made to containers, and possibly “transplant” them (e.g. when a new release of Ubuntu comes out, run your application on that new release without rebuilding it.) It sounds like a good idea at first! We stored images in Mercurial repositories, using the metashelf extension to track special files and permissions. This means that images operations were slow. Furthermore, it turns out that you can’t “rebase” a filesystem image like you would rebase a bunch of source commits. It kind of works as long as you’re only changing configuration files; but it’s useless if these configuration files are generated fromtemplates anyway. And it doesn’t work at all for binaries or bytecode.
As a result, authoring images was a slow, bulky process, requiring some extra tooling to be done efficiently. One of us (Louis Opter if I remember correctly) was generally in charge of updating the dotCloud official images; and everybody else hoped that they’d never have to do it.
That’s why Docker just used AUFS layers as-is. It was good enough, it was fast, and since an AUFS layer is just a bunch of files masking their counterparts in the original image, it means that you can still get the list of modifications very easily.
Dependency on AUFS
The dotCloud platform ran for about 5 years, and used AUFS all along. I’m going to be brutally honest: no other option would have worked for us. BTRFS used too much memory (and still does), because multiple containers running the same image lead to page cache duplication. Device Mapper thin provisioning didn’t exist (and has the same memory issues anyway). Ditto for ZFS. The other union filesystems (UnionMount, anyone?) were hardly maintained, and had tons of edge cases.
That’s why we used AUFS. It had its quirks, but for our use case, it worked beautifully. It allowed us to pack hundreds of containers on instances with 32 GB of RAM. It ran flawlessly everything we threw at it, except MongoDB (something to do with fancy mmap semantics), which prompted us to introduce volumes.
When we rolled out the first versions of Docker, we knew that the dependency on AUFS would eventually become an issue. We were particularly lucky, in the sense that it became an issue after Docker got enough traction to convince Red Hat to do a lot of the hard work involved to modularize Docker’s “graph driver.” That’s how Alexander Larsson and Vincent Batts ended up bringing support for BTRFS and Device Mapper to Docker.
Entering a running container
Back in the day, you couldn’t do the equivalent of docker exec or nsenter , because they both rely on the setns() system call. That system call appeared in Linux 3.0, in
- So how did we do, then?
If you have used LXC, you might remember lxc-attach , which gives you a console on a running container. It could have worked, but we found it rather capricious. It was acceptable if you just wanted to get a terminal in a runaway container; but you couldn’t depend on it as a remote command execution engine to setup database replication, for instance. It was conceptually closer to a serial console.
This leaves you with two options:
- patch your kernel to add support for the setns() system call;
- run an SSH server in your containers.
We did both. We had an abstract execution engine that would use setns() when available, and fallback to SSH otherwise.
This means that our containers were all running an SSH server. I was a huge fan of this SSH server, by the way, because it allowed me to do all kinds of cool hacks. This may come as a surprise, especially when one knows that I wrote this blog post, but that merely demonstrates my ability to change my mind, amirite?
Why have both? Because we wanted the performance and convenience of setns() ,but we didn’t want to rely on it and be forced to stick to an older kernel if a wild kernel vulnerability appeared.
The container daemon itself
Since containers are managed by LXC, you don’t need a long-running daemon (and at this point there was no container engine per se). In fact, if you scratch the surface, you realize that each container has its own long-running daemon: it’s lxc-start (it’ssimilar to rkt or runc ) and you connect to it using an abstract socket (from memory,@/var/lib/lxc/<containername> ).
This is great, because it’s simple. At least, it seems simple. Each container was fully contained (so to speak) within /var/lib/dotcloud/<containername> , so you could move a container simply by copying that directory to another machine. Of course, copying this directory while the container is running requires extra precautions; but there was something satisfying and UNIX-y in the fact that a container was just a directory, after all.
Of course we couldn’t help but build our own RPC layer
Perfect, we have our dc tool on our container nodes; now we need to slap an API on top of that to orchestrate deployments from a central place. Since containers are standalone, the process exposing that API doesn’t have to be bullet-proof, and you can update/upgrade/restart it without being worried about your containers being restarted.
Almost all the communication between processes and hosts was done using ZeroRPC. ZeroRPC is basically RPC over ZeroMQ, using MessagePack to serialize parameters, return values, and exceptions. MessagePack is similar to JSON, but way more efficient. (We didn’t care much about efficiency except for the high-traffic use cases like metrics and logs.)
If you’re curious about ZeroRPC, I presented it at PyCon a few years ago. Unfortunately, my French accent was a few orders of magnitude thicker than it is today (which says a lot) so you might struggle to understand me, sorry ☹
ZeroRPC allowed us to expose almost any Python module or class like this:
ZeroRPC also supports some fan-out topologies, including broadcast (all nodes receiving the function call; return value is discarded) and worker queue (all nodes subscribe to a “hub;” you send a function call to the hub, one idle worker will get it, so you get transparent load balancing of requests).
The original ZeroRPC was synchronous, but François-Xavier Bourlet implemented an asynchronous version (making use of coroutines), as well as “streaming” — basically, the ability for a function to return an iterator/generator, very useful for logs and live metrics! Andrea Luzzardi also implemented the zerotracer, which allowed us to get full traces of API calls using transparent middlewares. But I digress.
Let’s sprinkle micro-services all over
So here we are, with a “container” service running on each node, letting us do the following operations from a central place:
- create containers
- start/stop/destroy them
Listing containers (and gathering core host metrics) relied on a separate service called “hostinfo.” This service would just scan all the containers deployed locally, aggregate their satus, and send it all to a central place.
So thanks to “hostinfo” we can also list all containers from that central place. Cool.
In the very first versions, dotCloud was building your apps “in place,” i.e. when you push your code, the code would be copied to a temporary directory in the container (while it’s still running the previous version of your app!), the build would happen, then a switcheroo happens (a symlink is updated to point to the new version) and processes are restarted.
To keep things clean and simple, this build system was managed by a separate service, that directly accessed the container data structures on disk. So we had the “container manager,” “hostinfo,” and the “build manager,” all accessing a bunch of containers and configuration files in the same directory ( /var/lib/dotcloud ).
Then we added support for separate builds (probably similar to Heroku’s “slugs”). The build would happen in a separate container; then that container image would be transferred to the right host, and a switcheroo would happen (the old container is replaced by the new one).
We had the equivalent of volumes, so by making sure that the old and newcontainers were on the same host, this process could be used for stateful apps as well. This, by the way, was probably a Very Bad Idea; as ditching away stateful apps would have simplified things immensely for us. Keep in mind, though, that we were running not only web apps but also databases like MySQL, PostgreSQL, MongoDB, Redis, etc. I was one of the strong proponents of keeping stateful containers on board, and on retrospect I was very certainly wrong, since it made our lives way more complicated than they could have been. But I digress again!
Who would have guessed that sharing state through on-disk files was a bad idea?
To keep things simple and reduce impact to existing systems (at this point, we had a bunch of customers that each already generated more than $1K of monthly revenue, and we wanted to play safe), when we rolled out that new system, it was managed by another service. So now on our hosts we had the “container manager,” “hostinfo,” the “build manager” (for in-place builds), and the “deploy manager.”
(Small parenthesis: we didn’t transfer full container images, of course. We transferred only the AUFS rw layer; so that’s the equivalent of a two-line Dockerfile doing FROM python-nginx-uwsgi and RUN dotcloud-build.sh then pushing the resulting image around.)
Then we added a few extra services also accessing container data; in no specific order, there was a remote execution manager (used e.g. by the MySQL replication system), a metrics collector, and a bunch of hacks to work around EC2/EBS issues, kernel issues, out of memory killer, etc.; for instance in some scenarios, the OOM killer would leave the container in a weird state and we would need a few special operations to clean it up. In the early day this was manual ops work, but as soon as we had enough data it was automated.
The process tree on a container node looked like this:
So at this point we have a bunch of services accessing a bunch of on-disk structures. Locking was key. The problem is, that some operations are slow, so you don’t want to lock when unnecessary (e.g. you don’t want to lock everything while you’re merely pulling an image). Some operations can fail gracefully (e.g. it’s OK if metrics collection fails for a few minutes). Some operations are really important and you absolutely want to know if they went wrong (e.g. the stuff that watches over MySQL replica provisioning). Sometimes it’s OK to ignore a container for a bit (e.g. for metrics) but sometimes you absolutely want to know if it’s there (because if it’s not, a failover mechanism will spin it up somewhere else; so having containers disappearing in a transient manner would be bad).
To spice things further up, our ops toolkit was based on the dc CLI tool, so that tool had to play nice with everything else.
Still with us? Get ready for another episode of “embarrassing early start-up decisions.”
Dalek says, ORCHESTRATE
When your container platform runs 10 containers on a handful of nodes, you can place them manually; especially if you don’t create or resize containers all the time. But when you have thousands of containers (dotCloud peaked above 100,000 containers) running across hundreds of nodes, and your users constantly deploy and scale services, you need an orchestrator. More specifically, you need to automate resource scheduling.
In dotCloud’s case, we wanted to be able to make an API call to create a container, with the following parameters:
- some unique identifier for this container (combination of owner, app, service…)
- container base image and some templating parameters
- resources needed (mainly RAM)
- a “high availability token” which was used to prevent two containers with the
same token from running on the same host (e.g. the leader and follower of a replicated database) The API call should pick a machine to run the container, while honoring the various constraints specified in the call (available resources and HA token).
And by “orchestration” I mean “scheduling”
In theory, any good CS grad student will tell you that this seems like a perfectly good case to use some bin packing algorithm.
In practice, anybody who has worn a pager long enough knows that network latency and packet loss are both non-zero quantities, and that therefore, we are facing a distributed systems problem (aka potential nuclear waste dumpster fire).
Most “standard” algorithms assume that you know the full state of the cluster when taking a scheduling decision. But in our scenario, you don’t know the state of the cluster. You have to query each machine. The request has to go over the network, and then the machine has to read the state of all its containers before replying. Both operations (network round-trip and gathering container state) can and will take some time. Using aggressive timeouts (to avoid waiting forever for unreachable nodes) gets problematic when a host is very loaded (and takes a while to gather container state).
“Caching,” I heard someone say. Excellent idea! Caching is easy. Cache invalidation, however, is one of the hardest things in computer science. How, why, when would we have to invalidate the cache? Whenever some other system (other than the scheduler) makes changes to the containers. We ended up having lots of cluster maintenance tasks to move containers around, resize them, regroup them (if you regroup containers using the same image, you realize huge memory savings). These operations were implemented by relatively simple scripts, relying on the fact that each container was fully contained in a directory. To preserve these semantics, you need to somehow watch all your container configurations for changes, and trigger cache invalidation events when changes happen. Alternative option: implement these operations with the scheduler. That was not realistic. We were constantly expecting the unexpected (AWS “degraded performance” and “elevated error rates,” some customer spiking 100x, 1 Mpps distributed denial-of-service attacks, etc.) so we needed the ability to cobble solutions fast, without breaking too many things. Central scheduler was out, at least in the beginning.
Distributed, robust, suboptimal scheduling
Our scheduler would just broadcast the request to a subset of nodes, and place it in a retry queue. These nodes would try to acquire a lock (implemented by a centralized Redis). The one acquiring the lock would carry on and deploy your container. Once the container is up and running, another system takes care of removing it from the retry queue (which gets re-broadcasted once in a while).
That’s an extremely naive algorithm, but it’s also very resilient. By the way, some of the fastest search algorithms work by scattering your request on multiple nodes and
gathering only the first (fastest) replies, to make sure that you get a good response time (at the expense of correctness if some nodes are overloaded or down).
The main single point of failure was the Redis used for locking, and even that one was not a big deal since it didn’t really store anything. (We always meant to replace it with Zookeeper or Doozer, but it never turned out to be worth the pain!)
If this “algorithm” makes you cringe, that’s fair. We weren’t particularly proud of it, and we wanted our next container engine to support better semantics.
At this point, we really dreamed of a single point of entry to the container engine, to avoid locking issues. At the very least, all container metadata should be mediated by an engine exposing a clean API. We had a pretty good idea of what was needed, and that’s what shaped the first versions of the Docker API.
The first versions of Docker were still relying on LXC. The process tree on the container nodes would have looked like this:
“Waitaminute,” you say, “that’s exactly the same thing as before!”
Yes! But now, all our management processes (container, hostinfo, etc.) would go through “docker” instead of accessing container metadata on disk. No more crazy locking, no more hoping that everybody used the locking primitives correctly instead of accessing stuff directly, etc.
From LXC to libcontainer
Then, as containers picked up steam, LXC development (which was pretty much dead, or at least making very slow progress) came to life, and in a few months, there were more LXC versions than in the few years before. This broke Docker a few times, and that’s what led to the development of libcontainer, allowing to directly program cgroups and namespaces without going through LXC. You could put container processes directly under the container engine, but having an intermediary process helps a lot, so that’s what we did; it was named dockerinit .
The process tree now looked like this:
But now you have a problem: if the docker process is restarted, you end up orphaning all your “dockerinits.” For simplicity, docker and dockerinit share a bunch of file descriptors (giving access to the container’s stdout and stderr). The idea was to eventually make dockerinit a full-blown, standalone mini-daemon, allowing to pass FDs around across UNIX sockets, buffering logs, whatever’s needed.
Having a daemon to manage the containers (we’re talking low-level management here, i.e. listing, starting, getting basic metrics) is crucial. I’m sorry if I failed to convince you that it was important; but believe me, you don’t want to operate containers at scale without some kind of API. (Executing commands over SSH is fine until you have more than 10 containers per machine, then you really want a true API ☺)
When the daemon does too much
But at the same time, the Docker Engine has lots of features and complexity: builds, image management, semantic REST API over HTTP, etc.; those features are essential (they are what helped to drive container adoption, while vserver, openvz, jails, zones, LXC, etc. kept containers contained (sorry!) to the hosting world) but it’s totally reasonable that you don’t want all that code near your production stuff. That’s why in Docker Engine 1.11, we decided to break away all the low-level container management functions to containerd, while the rest would stay in the Docker Engine.
So the current solution is to delegate all the low-level management to containerd, and keep the rest in the Docker Engine.
You can think of containerd like a simplified Docker Engine. You can do the equivalent of docker ps , docker run , docker kill ; but it doesn’t deal with builds, and its API uses grpc instead of REST.
The process tree looks like this:
The big upside (which doesn’t appear on the diagram) is that the link between docker and containerd can be severed and reestablished, e.g. to restart or upgrade the Docker Engine. (This can be achieved with the live restore configuration option.)
Going full cycle
The dotCloud container engine started as a simple, standalone CLI tool. It was augmented with a collection of “sidekick” daemons, each providing a little bit of extra functionality. Eventually, this architecture showed its limits. The first version of the Docker Engine gathered all the features that were deemed necessary in a single daemon. Too many features? I don’t think so; precisely because these features made the success of Docker. The first versions of Docker sacrificed modularity, but that was only temporary. Over time, features were separated from the Docker Engine again. Today, you can use runc or containerd to run containers without the Docker Engine. Clustering features are provided by SwarmKit. External image builders are available, e.g. dockramp or box.
Two years ago, Docker committed to the motto “batteries included, but swappable.” It’s still doing exactly that: providing what most people need to build, ship, and run containerized apps, but giving an increasing number of options to remove whatever you don’t need or don’t like. And it all started with some really, really embarrassing container management code in Python, almost 10 years ago!