Docker at Spotify | Hacker News Reader

Docker at Spotify

continuousdelivery.uglyduckling.nl

65 pointspini4220 comments

I've been watching Docker ever since Flynn was announced, and started using it seriously this last December. Primarily, I'm trying to use it as a faster alternative to Chef/Puppet, running it over a Vagrant precise64 box, to potentially be something I can shove onto a foreign machine without much worry.

Overall, this has actually worked fairly well- getting certain containers up and running, and having them work together, hasn't been very difficult. But, there's one place that's consistently been a thorn in my side: PostgreSQL.

Currently, there is this article in Docker.io's docs[1] that claims to help you set up a Postgres container, but so far I've not seen it work for me or anyone else I've spoken to about it. It's primarily come from how Postgres is traditionally installed/ran:

1. Service is started. This can either be an 'active' Postgres server, or in the background. This creates a Postmaster.pid file, which ensures no other Postgres servers are running on the machine(I'm not entirely sure why, beyond ease-of-use scenarios).

2. Service is stopped. This deletes the Postmaster.pid file.

In the case of Docker containers, this presents a bit of a problem. If you try to start the Postgres container that way, it will run, but the Postmaster.pid seems to sit outside of it. In the process, once you shut that container down, you actually can't start another without error- the Postmaster.pid stops you.

The other alternative, is having a container running with Postgres as a background process to something else. This can work, but you start having to use services that can keep a process running while letting you access /bin/bash to still get at Postgres.

Neither of these potential solutions allow you to retain data, either.

I've done a lot of poking around, but I haven't had a lot of success in figuring out how to get past this. Those who are using Docker as a Puppet/Chef replacement, especially with Vagrant- what are you usually doing for a database solution?

1. http://docs.docker.io/en/latest/examples/postgresql_service/

Here's my docker file for running a postgres 9.3 with plv8

    FROM    boothead/saucy

    ENV     DEBIAN_FRONTEND noninteractive

    # Add repository and install PostgreSQL 9.3
    RUN     echo "deb http://apt.postgresql.org/pub/repos/apt/ precise-pgdg main 9.3" >> /etc/apt/sources.list
    RUN     apt-get update
    RUN     apt-get -y --force-yes -f install postgresql-9.3 postgresql-client-9.3 libicu48 wget

    RUN     wget http://www.mirrorservice.org/sites/archive.ubuntu.com/ubuntu//pool/universe/libv/libv8/libv8-3.7.12.22_3.7.12.22-3_amd64.deb
    RUN     dpkg -i libv8-3.7.12.22_3.7.12.22-3_amd64.deb
    RUN     apt-get -y --force-yes -f install postgresql-9.3-plv8


    # Postgres is started now, shut it down and replace config
    RUN     service postgresql stop
    ADD     pg_hba.conf     /etc/postgresql/9.3/main/
    ADD     pg_ident.conf   /etc/postgresql/9.3/main/
    ADD     postgresql.conf /etc/postgresql/9.3/main/

    # Create superuser, tempo db and plv8 language
    ADD     create_user.sh /
    RUN     /bin/bash /create_user.sh
    RUN     rm /create_user.sh

    # Configure for running in container
    EXPOSE  5432
    # VOLUME ["/var/lib/postgresql/9.3/main"]
    CMD     ["/bin/su", "postgres", "-c", "/usr/lib/postgresql/9.3/bin/postgres -D /var/lib/postgresql/9.3/main -c config_file=/etc/postgresql/9.3/main/postgresql.conf"]

create_user.sh is basically service postgresql start; su echo <SQL> | psql; service postgresql stop;

This is fine for me to spin up a postgres instance at some known state for development, but it's a bit dissatisfying for managing the data separately from the container. Anyone got any tips for integrating volumes with this?

I've been trying out Docker recently but haven't tried running persistent services in it yet (only long running programs like apps). From what I've gathered so far, Docker allows you to specify volumes (via -v /host/path:/container/path) that are mounted directly and don't get included in the container's copy-on-write filesystem (basically shared directories with the host). You can use this to pass data into the container (ex: a SQL file to init your schema) or pass data out of the container (ex: the DB files themselves or the logs). I've been using the latter to centralize Dokku logs[1].

To split out the Postgres setup and the initial schema setup, one idea might be to have a the CMD for the container run a script to automatically check for a bootstrap file on startup and run it. The bootstrap file could be specified via a volume mount. To use the container with a different dev setup (i.e. a different bootstrap SQL script) you'd simply start up the container with the shared directory pointing somewhere else.

If you want to save the Postgres data files themselves outside of the container you can again do it with volume mounts but you'll need some way to keep track of which goes where. The volume mounts are specified each time you startup the container so you need something to save those. For Dokku specifically there's the PG plugin and it looks like it does exactly this[2]. I haven't use it but I guess it validates the idea.

This seems like a general trend with Docker; it's really cool tech but it's pretty low level so you need something atop it to make usage smoother.

[1]: https://github.com/sehrope/dokku-logging-supervisord

[2]: https://github.com/Kloadut/dokku-pg-plugin/blob/master/comma...

Maybe you are not shutting Postgres down correctly? 'docker kill' is kill `kill -9`, it doesn't let the container shutdown gracefully. `docker stop` on the hand sends SIGTERM to the container and gives it a chance to shut down correctly.

As for running Postgres as a background process of something else: I actually recommend this approach. I think running a single process inside a Docker container is a bad idea. By doing that, it does not start important system services such as syslog, and does not allow the container to really work correctly as a proper Unix system should. That is why we have created https://github.com/phusion/baseimage-docker, and we base all our Docker containers on that. The Github page that I linked to explains the problem in detail, explains why a lot of Docker containers are actually not working properly, explains how it solves that problem, and explains how you can use the solution.

Have you looked at https://github.com/jeffutter/dokku-postgresql-plugin ?

It's built to run with dokku, but the setup steps in the install and command file show how to setup postgres successfully.

I like how no one is paying heed to the "Don't use Docker in production" warning on the Docker site. Every once in a while they release a version that's buggy (like 0.7.3) and we fix to the previous version and wait for the next version.

They let devs ssh onto production boxes and do what they want?!?! "Things diverge pretty quickly" - I'm sure they do. That sounds like a recipe for disaster...

Docker looks promising but until there's a way to allocate maximum resources to a container I wouldn't use it in production. VMs are much slower to start but at least a runaway process on one won't affect other VMs on the same hardware.

Docker only lets you hint that only a certain number of cores should be used, so a bad process might monopolise the physical machine. Also until other tools (ansible, salt, chef, etc) help with provisioning you need to edit the dockerfile to change parameters (eg for settings that depend on your environment like smtp end points test vs prod, etc).

If these points are addressed I think it'll be awesome though.

Docker lets you restrict memory and cpu consumption on a per-container basis. Configuration like endpoints are usually fed into the container via environment variables, not by changing stuff around in the Dockerfile. This way you can use the same container in multiple environments.

It's fairly easy to integrate configuration management tools into your Docker workflow - I wrote a blog post about it here: http://kartar.net/2013/12/building-puppet-apps-inside-docker....

Aren't cgroups be more then a hitting ? Or do you mean that by default does not setup strict enough cgroups ?

It seems cpu.shares allow a proportional amount of CPU to be allocated, but LXC may use up to 100% of all idle CPUs. You can't specify a hard-limit of "only use 1GHz when I have a 2GHz CPU" apparently (according to here: http://comments.gmane.org/gmane.linux.kernel.containers.lxc....). So I guess that could mean that if the containers are using 100% of all CPUs, there wouldn't be resources left e.g. for monitoring on the host? Anyone got any experience with this?

Take a look at packer.io

Thanks. I had seen packer.io but then forgot about it.

I had another look at Docker in November and had issues installing what I needed via apt for some reason. It wasn't a drop-in replacement for VMs for several reasons (I wanted an alternative backend to provision with salt), but when it hits 1.0 I'll have a look again, if not before.

Definitely it is not. You'd have to prepare the docker instance a little bit differently than you would for VM.

For example: Docker does not run upstart by default.

re: issues with installing via apt => Docker public registry provided a stock Ubuntu 12.04 with a lot of things stripped out including the /etc/apt/sources.list (it contains only 1 entry). You'd have to populate that sources.list with a Ubuntu/Debian that you already have.

That's just by default though. You can easily run upstart (or other init styles) inside Docker.

Additionally you can use a number of base images if the stock Ubuntu image doesn't suit you.

If there are any Spotify people reading, please please please fix the search/filter regression. You broke it a year-ish ago. It's nice to hear about innovations but a bit galling to paying customers when you break an essential feature and don't fix it.

To reproduce: be on an artist page. Now try and find a track with a given name on that page. I dare you. Without reading every single track with your eyes. You used to be able to filter that list.

http://community.spotify.com/t5/Spotify-Ideas/0-8-8-Bring-ba...

"* Please note Docker is currently under heavy developement. It should not be used in production (yet)." [1]

[1] http://www.docker.io/learn_more/

This is addressed in the talk. It's a short talk, so may be worth listening to it to see how it was used.

Agreed, but I'd have thought that it was worth raising for those who are not familiar with where Docker is in the development cycle.

Why Docker? Why not Packer.io? Why No Love for OSv?

Reply on news.ycombinator.com