Django UWSGI deploy demo

This repository demonstrates the principles behind the talk “High-Availability Django” given by Frankie Dintino of The Atlantic at the 2016 Djangocon.

In order to simulate the deployment of a large-ish Django project, it uses the source for addons.mozilla.org, found here.

Running

In order to run this demo VM, Docker, Fabric, and VirtualBox must be installed. Once installed, the VM can be provisioned by running (from the root of the checked out repository):

git submodule init
git submodule update
# Skip the next two steps if you don't want docker-machine
docker-machine create --driver virtualbox addons
eval $(docker-machine env addons)
docker-compose up -d
fab init

During the fab init step you will be prompted once when the database is initialized, and then again to create the superuser.

Using docker-machine is not required unless you are using the older Docker Toolbox on Mac (or you have installed it with brew install docker). Keep in mind however that by default, running docker-compose up without using a docker-machine VM will bind to port 80 on your localhost, which means it will need to be unused, so you may wish to use docker-machine even if you don't need to. Alternatively, if you are using Docker for Mac, you can use https://github.com/nlf/dlite to have docker use a separate bridged network (and thus separate ips for its containers).

Grab the ip address using the command fab ip (unless you are not using docker-machine or dlite, in which case it will be 127.0.0.1), and then add to your /etc/hosts files:

192.168.64.7  live.addons stage.addons

Replacing 192.168.64.7 with the ip address for docker.

After running fab init, you should be able to visit http://live.addons/ and http://stage.addons/ to see the two initial builds (note that there may be a delay while the code initializes). You can view a live-updating uwsgi status dashboard at http://live.addons/uwsgi/

Creating new builds

Builds are stored in folders, as part of the sequence {a, b, c, ..., z, aa, ab, ...}

In many scenarios, builds will cycle between a, b, and c, and the unused build (the one linked neither to live or stage) can be removed or archived after each build.

To find the next unused build folder, run fab find_free_build.

Then, to create the build, run, e.g. fab create_build:z where “z” should be whatever build name was returned from find_free_build.

Staging and deploying builds

To set a build as active (which will create a wsgi configuration that spawns vassal in uWSGI emperor mode), run fab activate:z.

To then stage this build, run fab stage:z (again, the string after the “:” being the build name), which will unlink the current stage build and link the specified build to stage.

To “warm up” a build by hitting it with concurrent requests, use the command fab warmup:z (the build name is optional and defaults to the stage build).

To swap the stage and live builds, run fab swap_live. By default, this will swap live with the stage build, but it is possible to specify a different build by, e.g. fab swap_live:z.

To spin down an old build (it cannot be the current live or stage builds), run fab deactivate:z.

fab init and fab create_build

(Note: This repository's directory will be mounted as /code within the docker container, so when you see paths like /code/deploy/... below, keep in mind that this refers also to ./deploy/... within this repository.)

Running fab init executes a number of steps to get a stage and live build up and running:

• Builds all of the requirements into wheels, output to /code/deploy/deps/wheelhouse, which then allows for faster pip install when creating new addons-server builds

• Creates two builds, “a” and “b”, in directories /code/deploy/builds/{a,b} (fab create_build:a and fab create_build:b). The create_build command does the following:

  • Creates symlinks for the static and media directories of the build so that nginx static serving works correctly.

    ^{Note: All static assets (static files and user-uploaded media) live outside of the builds directory, in /code/deploy/assets. This is to allow for the common use case where static files and user media use a distinct mount or partition from the rest of the code. In order to keep all static file urls distinct (beyond whatever staticfile cache busting is used), the static files for all builds are kept in separate directories, e.g. /code/deploy/assets/static/a.}

  • Uses rsync to copy the source in the addons-server submodule to the build directory, and copies the contents of /code/src on top of that.

  • From the build directory, runs git init and commits a randomly generated file. Mozilla’s addons-server uses the git commit hash to cache-bust minified assets, so this ensures that this step won’t fail and that the hash will be unique for each build.

  • Runs npm install as well as pip install with the requirements files, using the --no-index flag and our wheeldir of code/deploy/deps/wheelhouse to keep this speedy.

  • Runs manage.py collectstatic and manage.py compress_assets (the latter being a jingo-minify management command specific to the mozilla addons-server project) from the build’s virtualenv.

  • Determines what port number corresponds to the current build name, and creates /code/deploy/builds/a/live.conf and /code/deploy/builds/a/stage.conf, which are nginx configuration files that specify the upstream. This will only come into play if the build gets designated as the stage or live build (since nginx includes /code/deploy/builds/_live/live.conf and /code/deploy/builds/_stage/stage.conf).

  • Creates symlinks to the uwsgi skeleton file /code/conf/uwsgi/vassal.skel at /code/deploy/builds/{a,b}/vassal.ini. This step causes the uWSGI emperor to spawn a vassal that will manage the zerg instances. At this point the application on the web container will be listening on ports 2000 and 2010 (for the a and b builds, respectively), but because there are no zerg instances attached to the vassal http requests will not yet route.

• After the builds “a” and “b” are created and linked, respectively, to /code/deploy/builds/_live and /code/deploy/builds/_stage, fab init initiates the database and populates it with sample test data, executed using the manage.py in the _live (read “a”) build’s virtualenv.

• It then symlinks to the uwsgi skeleton file /code/conf/uwsgi/zerg.skel from /code/deploy/builds/{a,b}/zerg.ini. This “activates” (fab activate:a) the build, spawning zerg instances that attach to the vassal.

• Lastly, nginx -s reload is run on the nginx container, updating the upstreams.

Troubleshooting

• Problem: You encounter the error “Error response from daemon: client is newer than server (client API version: 1.24, server API version: 1.22)”
• Solution: export DOCKER_API_VERSION=1.22