Zulip uses the standard Django system for doing schema migrations. There is some example usage in the new feature tutorial.
This page documents some important issues related to writing schema migrations.
If your database migration is just to reflect new fields in
models.py, you’ll want to just run
manage.py migrationsto generate a migration file (remember to
git addit). If the
models.pychange is a single field or table being added, the automatically generated names are nice andx clear, but if the migration changes more than one thing, you’ll get a name determined by the date like
0089_auto_20170710_1353.pyand will want to rename the migration file.
Naming: Please provide clear names for new database migrations (e.g.
0072_realmauditlog_add_index_event_time.py). Since in the Django migrations system, the filename is the name for the migration, this just means moving the migration file to have a reasonable name. Note that
tools/test-migrationswill fail in Travis CI if a migration has bad name of the form
0089_auto_20170710_1353.py, which are what Django generates automatically for nontrivial database schema changes.
Large tables: For large tables like Message and UserMessage, you want to take precautions when adding columns to the table, performing data backfills, or building indexes. We have a
zerver/lib/migrate.pylibrary to help with adding columns and backfilling data. For building indexes on these tables, we should do this using SQL with postgres’s CONCURRENTLY keyword.
Numbering conflicts across branches: If you’ve done your schema change in a branch, and meanwhile another schema change has taken place, Django will now have two migrations with the same number. There are two easy way to fix this:
If your migrations were automatically generated using
manage.py makemigrations, a good option is to just remove your migration and rerun the command after rebasing. Remember to
git rebaseto do this in the the commit that changed
models.pyif you have a multi-commit branch.
If you wrote code as part of preparing your migrations, or prefer this workflow, you can use run
./tools/renumber-migrations, which renumbers your migration(s) and fixes up the “dependencies” entries in your migration(s). The tool could use a bit of work to prompt unnecessarily less, but it will update the working tree for you automatically (you still need to do all the
git addcommands, etc.).
Atomicity. By default, each Django migration is run atomically inside a transaction. This can be problematic if one wants to do something in a migration that touches a lot of data and would best be done in batches of e.g. 1000 objects (e.g. a
UserMessagetable change). There is a new Django feature added in Django 1.10 that makes it possible to add
atomic=Falseat the top of a
Migrationclass and thus not have the entire migration in a transaction. This should make it possible to use the batch update tools in
zerver/lib/migrate.py(originally written to work with South) for doing larger database migrations.
Accessing code and models in RunPython migrations. When writing a migration that includes custom python code (aka
RunPython), you almost never want to import code from
zerveror anywhere else in the codebase. If you imagine the process of upgrading a Zulip server, it goes as follows: first a server admin checks out a recent version of the code, and then runs any migrations that were added between the last time they upgraded and the current check out. Note that for each migration, this means the migration is run using the code in the server admin’s check out, and not the code that was there at the time the migration was written. This can be a difference of thousands of commits for installations that are only upgraded occasionally. It is hard to reason about the effect of a code change on a migration that imported it so long ago, so we recommend just copying any code you’re tempted to import into the migration file directly, and have a linter rule enforcing this.
There is one special case where this doesn’t work: you can’t copy the definition of a model (like
Realm) into a migration, and you can’t import it from
zerver.modelsfor the reasons above. In this situation you should use Django’s
apps.get_modelto get access to a model as it is at the time of a migration. Note that this will work for doing something like
Realm.objects.filter(..), but shouldn’t be used for accessing
Realm.subdomainor anything not related to the Django ORM.
Making large migrations work. Major migrations should have a few properties:
Unit tests. You’ll want to carefully test these, so you might as well write some unit tests to verify the migration works correctly, rather than doing everything by hand. This often saves a lot of time in re-testing the migration process as we make adjustments to the plan.
Run in batches. Updating more than 1K-10K rows (depending on type) in a single transaction can lock up a database. It’s best to do lots of small batches, potentially with a brief sleep in between, so that we don’t block other operations from finishing.
Rerunnability/idempotency. Good migrations are ones where if operational concerns (e.g. it taking down the Zulip server for users) interfere with it finishing, it’s easy to restart the migration without doing a bunch of hand investigation. Ideally, the migration can even continue where it left off, without needing to redo work.
Multi-step migrations. For really big migrations, one wants to split the transition into into several commits that are each individually correct, and can each be deployed independently:
First, do a migration to add the new column to the Message table and start writing to that column (but don’t use it for anything)
Second, do a migration to copy values from the old column to the new column, to ensure that the two data stores agree.
Third, a commit that stops writing to the old field.
Any cleanup work, e.g. if the old field were a column, we’d do a migration to remove it entirely here.
This multi-step process is how most migrations on large database tables are done in large-scale systems, since it ensures that the system can continue running happily during the migration.
Automated testing for migrations¶
Zulip has support for writing automated tests for your database
migrations, using the
MigrationsTestCase test class. This system is
inspired by a great blog post on
We have integrated this system with our test framework so that if you
use_db_models decorator, you can use some helper methods
test_classes.py and friends from inside the tests (which is
normally not possible in Django’s migrations framework).
If you find yourself writing logic in a
RunPython migration, we
highly recommend adding a test using this framework. We may end up
deleting the test later (they can get slow once they are many
migrations away from current), but it can help prevent disaster where
an incorrect migration messes up a database in a way that’s impossible
to undo without going to backups.