Glyph Lefkowitz: What Am Container

Perhaps you are a software developer.

Perhaps, as a developer, you have recently become familiar with the term "containers".

Perhaps you have heard containers described as something like "LXC, but better", "an application-level interface to cgroups" or "like virtual machines, but lightweight", or perhaps (even less usefully), a function call. You've probably heard of "docker"; do you wonder whether a container is the same as, different from, or part of an Docker?

Are you are bewildered by the blisteringly fast-paced world of "containers"? Maybe you have no trouble understanding what they are - in fact you might be familiar with a half a dozen orchestration systems and container runtimes already - but frustrated because this seems like a whole lot of work and you just don't see what the point of it all is?

If so, this article is for you.

I'd like to lay out what exactly the point of "containers" are, why people are so excited about them, what makes the ecosystem around them so confusing. Unlike my previous writing on the topic, I'm not going to assume you know anything about the ecosystem in general; just that you have a basic understanding of how UNIX-like operating systems separate processes, files, and networks.¹

At the dawn of time, a computer was a single-tasking machine. Somehow, you'd load your program into main memory, and then you'd turn it on; it would run the program, and (if you're lucky) spit out some output onto paper tape.

When a program running on such a computer looked around itself, it could "see" the core memory of the computer it was running on, any attached devices, including consoles, printers, teletypes, or (later) networking equipment. This was of course very powerful - the program had full control of everything attached to the computer - but also somewhat limiting.

This mode of addressing hardware is limiting because it meant that programs would break the instant you moved them to a new computer. They had to be re-written to accommodate new amounts and types of memory, new sizes and brands of storage, new types of networks. If the program had to contain within itself the full knowledge of every piece of hardware that it might ever interact with, it would be very expensive indeed.

Also, if all the resources of a computer were dedicated to one program, then you couldn't run a second program without stomping all over the first one - crashing it by mangling its structures in memory, deleting its data by overwriting its data on disk.

So, programmers cleverly devised a way of indirecting, or "virtualizing", access to hardware resources. Instead of a program simply addressing all the memory in the whole computer, it got its own little space where it could address its own memory - an address space, if you will. If a program wanted more memory, it would ask a supervising program - what we today call a "kernel" - to give it some more memory. This made programs much simpler: instead of memorizing the address offsets where a particular machine kept its memory, a program would simply begin by saying "hey operating system, give me some memory", and then it would access the memory in its own little virtual area.

In other words: memory allocation is just virtual RAM.

Virtualizing memory - i.e. ephemeral storage - wasn't enough; in order to save and transfer data, programs also had to virtualize disk - i.e. persistent storage. Whereas a whole-computer program would just seek to position 0 on the disk and start writing data to it however it pleased, a program writing to a virtualized disk - or, as we might call it today, a "file" - first needed to request a file from the operating system.

In other words: file systems are just virtual disks.

Networking was treated in a similar way. Rather than addressing the entire network connection at once, each program could allocate a little slice of the network - a "port". That way a program could, instead of consuming all network traffic destined for the entire machine, ask the operating system to just deliver it all the traffic for, say, port number seven.

In other words: listening ports are just virtual network cards.

Getting bored by all this obvious stuff yet? Good. One of the things that frustrates me the most about containers is that they are an incredibly obvious idea that is just a logical continuation of a trend that all programmers are intimately familiar with.

All of these different virtual resources exist for the same reason: as I said earlier, if two programs need the same resource to function properly, and they both try to use it without coordinating, they'll both break horribly.²

UNIX-like operating systems more or less virtualize RAM correctly. When one program grabs some RAM, nobody else - modulo super-powered administrative debugging tools - gets to use it without talking to that program. It's extremely clear which memory belongs to which process. If programs want to use shared memory, there is a very specific, opt-in protocol for doing so; it is basically impossible for it to happen by accident.

However, the abstractions we use for disks (filesystems) and network cards (listening ports and addresses) are significantly more limited. Every program on the computer sees the same file-system. The program itself, and the data the program stores, both live on the same file-system. Every program on the computer can see the same network information, can query everything about it, and can receive arbitrary connections. Permissions can remove certain parts of the filesystem from view (i.e. programs can opt-out) but it is far less clear which program "owns" certain parts of the filesystem; access must be carefully controlled, and sometimes mediated by administrators.

In particular, the way that UNIX manages filesystems creates an environment where "installing" a program requires manipulating state in the same place (the filesystem) where other programs might require different state. Popular package managers on UNIX-like systems (APT, RPM, and so on) rarely have a way to separate program installation even by convention, let alone by strict enforcement. If you want to do that, you have to re-compile the software with ./configure --prefix to hard-code a new location. And, fundamentally, this is why the package managers don't support installing to a different place: if the program can tell the difference between different installation locations, then it will, because its developers thought it should go in one place on the file system, and why not hard code it? It works on their machine.

In order to address this shortcoming of the UNIX process model, the concept of "virtualization" became popular. The idea of virtualization is simple: you write a program which emulates an entire computer, with its own storage media, network devices, and then you install an operating system on it. This completely resolves the over-sharing of resources: a process inside a virtual machine is in a very real sense running on a different computer than programs running on a different virtual machine on the same physical device.

However, virtualiztion is also an extremly heavy-weight blunt instrument. Since virtual machines are running operating systems designed for physical machines, they have tons of redundant hardware-management code; enormous amounts of operating system data which could be shared with the host, but since it's in the form of a disk image totally managed by the virtual machine's operating system, the host can't really peek inside to optimize anything. It also makes other kinds of intentional resource sharing very hard: any software to manage the host needs to be installed on the host, since if it is installed on the guest it won't have full access to the host's hardware.

I hate using the term "heavy-weight" when I'm talking about software - it's often bandied about as a content-free criticism - but the difference in overhead between running a virtual machine and a process is the difference between gigabytes and kilobytes; somewhere between 4-6 orders of magnitude. That's a huge difference.

This means that you need to treat virtual machines as multi-purpose, since one VM is too big to run just a single small program. Which means you often have to manage them almost as if they were physical harware.

When we run a program on a UNIX-like operating system, and by so running it, grant it its very own address space, we call the entity that we just created a "process".

This is how to understand a "container".

A "container" is what we get when we run a program and give it not just its own memory, but its own whole virtual filesystem and its own whole virtual network card.

The metaphor to processes isn't perfect, because a container can contain multiple processes with different memory spaces that share a single filesystem. But this is also where some of the "container ecosystem" fervor begins to creep in - this is why people interested in containers will religiously exhort you to treat a container as a single application, not to run multiple things inside it, not to SSH into it, and so on. This is because the whole point of containers is that they are lightweight - far closer in overhead to the size of a process than that of a virtual machine.

A process inside a container, if it queries the operating system, will see a computer where only it is running, where it owns the entire filesystem, and where any mounted disks were explicitly put there by the administrator who ran the container. In other words, if it wants to share data with another application, it has to be given the shared data; opt-in, not opt-out, the same way that memory-sharing is opt-in in a UNIX-like system.

So why is this so exciting?

In a sense, it really is just a lower-overhead way to run a virtual machine, as long as it shares the same kernel. That's not super exciting, by itself.

The reason that containers are more exciting than processes is the same reason that using a filesystem is more exciting than having to use a whole disk: sharing state always, inevitably, leads to brokenness. Opt-in is better than opt-out.

When you give a program a whole filesystem to itself, sharing any data explicitly, you eliminate even the possibility that some other program scribbling on a shared area of the filesystem might break it. You don't need package managers any more, only package installers; by removing the other functions of package managers (inventory, removal) they can be radically simplified, and less complexity means less brokenness.

When you give a program an entire network address to itself, exposing any ports explicitly, you eliminate even the possibility that some rogue program will expose a security hole by listening on a port you weren't expecting. You eliminate the possibility that it might clash with other programs on the same host, hard-coding the same port numbers or auto-discovering the same addresses.

In addition to the exciting things on the run-time side, containers - or rather, the things you run to get containers, "images"³, present some compelling improvements to the build-time side.

On Linux and Windows, building a software artifact for distribution to end-users can be quite challenging. It's challenging because it's not clear how to specify that you depend on certain other software being installed; it's not clear what to do if you have conflicting versions of that software that may not be the same as the versions already available on the user's computer. It's not clear where to put things on the filesystem. On Linux, this often just means getting all of your software from your operating system distributor.

You'll notice I said "Linux and Windows"; not the usual (linux, windows, mac) big-3 desktop platforms, and I didn't say anything about mobile OSes. That's because on macOS, Android, iOS, and Windows Metro, applications already run in their own containers. The rules of macOS containers are a bit weird, and very different from Docker containers, but if you have a Mac you can check out ~/Library/Containers to see the view of the world that the applications you're running can see. iOS looks much the same.

This is something that doesn't get discussed a lot in the container ecosystem, partially because everyone is developing technology at such a breakneck pace, but in many ways Linux server-side containerization is just a continuation of a trend that started on mainframe operating systems in the 1970s and has already been picked up in full force by mobile operating systems.

When one builds an image, one is building a picture of the entire filesystem that the container will see, so an image is a complete artifact. By contrast, a package for a Linux package manager is just a fragment of a program, leaving out all of its dependencies, to be integrated later. If an image runs on your machine, it will (except in some extremely unusual circumstances) run on the target machine, because everything it needs to run is fully included.

Because you build all the software an image requires into the image itself, there are some implications for server management. You no longer need to apply security updates to a machine - they get applied to one application at a time, and they get applied as a normal process of deploying new code. Since there's only one update process, which is "delete the old container, run a new one with a new image", updates can roll out much faster, because you can build an image, run tests for the image with the security updates applied, and be confident that it won't break anything. No more scheduling maintenance windows, or managing reboots (at least for security updates to applications and libraries; kernel updates are a different kettle of fish).

That's why it's exciting. So why's it all so confusing?⁵

Fundamentally the confusion is caused by there just being way too many tools. Why so many tools? Once you've accepted that your software should live in images, none of the old tools work any more. Almost every administrative, monitoring, or management tool for UNIX-like OSes depends intimately upon the ability to promiscuously share the entire filesystem with every other program running on it. Containers break these assumptions, and so new tools need to be built. Nobody really agrees on how those tools should work, and a wide variety of forces ranging from competitive pressure to personality conflicts make it difficult for the panoply of container vendors to collaborate perfectly⁴.

Many companies whose core business has nothing to do with infrastructure have gone through this reasoning process:

1. Containers are so much better than processes, we need to start using them right away, even if there's some tooling pain in adopting them.
2. The old tools don't work.
3. The new tools from the tool vendors aren't ready.
4. The new tools from the community don't work for our use-case.
5. Time to write our own tool, just for our use-case and nobody else's! (Which causes problem #3 for somebody else, of course...)

A less fundamental reason is too much focus on scale. If you're running a small-scale web application which has a stable user-base that you don't expect a lot of growth in, there are many great reasons to adopt containers as opposed to automating your operations; and in fact, if you keep things simple, the very fact that your software runs in a container might obviate the need for a system-management solution like Chef, Ansible, Puppet, or Salt. You should totally adopt them and try to ignore the more complex and involved parts of running an orchestration system.

However, containers are even more useful at significant scale, which means that companies which have significant scaling problems invest in containers heavily and write about them prolifically. Many guides and tutorials on containers assume that you expect to be running a multi-million-node cluster with fully automated continuous deployment, blue-green zero-downtime deploys, a 1000-person operations team. It's great if you've got all that stuff, but building each of those components is a non-trivial investment.

So, where does that leave you, my dear reader?

You should absolutely be adopting "container technology", which is to say, you should probably at least be using Docker to build your software. But there are other, radically different container systems - like Sandstorm - which might make sense for you, depending on what kind of services you create. And of course there's a huge ecosystem of other tools you might want to use; too many to mention, although I will shout out to my own employer's docker-as-a-service Carina, which delivered this blog post, among other things, to you.

You shouldn't feel as though you need to do containers absolutely "the right way", or that the value of containerization is derived from adopting every single tool that you can all at once. The value of containers comes from four very simple things:

1. It reduces the overhead and increases the performance of co-locating multiple applications on the same hardware,
2. It forces you to explicitly call out any shared state or required resources,
3. It creates a complete build pipeline that results in a software artifact that can be run without special installation or set-up instructions (at least, on the "software installation" side; you still might require configuration, of course), and
4. It gives you a way to test exactly what you're deploying.

These benefits can combine and interact in surprising and interesting ways, and can be enhanced with a wide and growing variety of tools. But underneath all the hype and the buzz, the very real benefit of containerization is basically just that it is fixing a very old design flaw in UNIX.

Containers let you share less state, and shared mutable state is the root of all evil.