Docker: First Impressions
Today I deployed my first Docker based application. I just wanted to get down some basic thoughts about how it went down etc. Nearly all the hosted (ie not turnkey) apps that we have at work have some form of git-based deployment strategy. (Aside: Don’t say something silly to yourself like “git is not a deployment strategy!” It totally is, you just don’t like the tradeoffs.) Each one has it’s own special snowflake of push vs. pull, scripted with perl vs scripted with shell, how or even if it automates database migrations; I could go on.
So I was tired of a million different ways to do it, so what did I do? I came up with another! Well, I didn’t come up with it obviously, but I decided I wanted to try Docker out as a deployment strategy. I’ve used Docker for testing for nearly a year at this point and have really enjoyed it, but I didn’t have a clue what it would be like for deployment.
Here are some of the things that I discovered:
Docker without the registry is really not like git. What I mean by this is that while there are layers inside the container (or is it layers in the image?) those layers are not used efficiently in a peer to peer setting. While a user can run
docker save myapp | ssh [email protected] docker load -and it will defintely work, it will push the entire thing instead of just the new layers. It would be great if docker worked more in this use case as it is nice when you are just getting started.
Setup was not magically easy. While it’s true that I was able to just build directly off the official Perl 5.20.1 image, that image is relatively large (which I’m working on). On top of that, installing the latest Perl version is trivial with perlbrew and plenv. What’s the most frustrating is the deps. Initially I just put a
RUN cpanm --installdeps .in the
Dockerfile. But that meant that if I ever needed to rebuild my image (hint: you will need to rebuild your image) it would need to install all the deps again. This app is super small, it literally has 10 direct deps, but it takes a long time even with the testing turned off. I decided to go with carton to handle the deps. It’s actually been pretty nice, but I’m not totally comfortable with it yet.
What about development? For an application that you actually work on, I don’t really see how Docker, or at least using the same Dockerfile, makes any sense at all. My Docker image (or is it the container?) has the actual project embedded inside of it; there’s no way it would make sense to use the same image for dev; you’d wanna use a volume for the actual code, but for deployment that would be terrible. At the very least there is FINALLY a plan to allow specifying a Dockerfile, so we could have a Dockerfile.live and a Dockerfile.dev.
LXC may be effectively metal, but Docker is a serious resource user. This is likely just me being silly, but I might as well put it down. When I provision a Linux server I usually starve it’s resources, if only to show my coworkers how much better Linux is in every way than Windows (which nearly everything at work runs on.) What that meant for this project was that it started with a meager 256M of RAM. While the OS itself takes something like 50M of RAM, and the app is something like 12M, at least while loading up the images, Docker spikes to 400M of RAM. At some point it triggered an OOM so bad that there was a kernel panic. I don’t really fault Docker for this in the least, but when creating a server that will host a single container, at least consider some basic extra stuff for the host.
Automation Rules! What I am the most excited about is that I created a Dockerfile that would automate the installation of the MS ODBC Driver. The only slow part of that is that testing DBI is slow, aside from that it’s really fast and so nice to have automated. I will likely publish it, though I don’t see how I could legally add it to the index.
Overall, I suspect that this will be a win in the long run, but only after I set up a private registry and at least do this on one other project (so that this is not it’s own special snowflake like all the rest.) I also think that setting up a development Dockerfile will be a huge win, once it’s an actual option.
In another post in the future I plan on contrasting Docker and Vagrant, since one of my coworkers really likes Vagrant. Not to actually go into detail there, but I think Vagrant is fine and definitely has it’s place, but there are a number of drawbacks that I think make Docker superior.Posted Wed, Jan 28, 2015