Schema migrations

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 typically want to just:

    • Rebase your branch before you start (this may save work later).

    • Update the model class definitions in zerver/models/*.py.

    • Run ./manage.py makemigrations to generate a migration file

    • Rename the migration file to have a descriptive name if Django generated used a date-based name like 0089_auto_20170710_1353.py (which happens when the changes are to multiple models and Django).

    • git add the new migration file

    • Run tools/provision to update your local database to apply the migrations.

    • Commit your changes.

  • For more complicated migrations where you need to run custom Python code as part of the migration, it’s best to read past migrations to understand how to write them well. git grep RunPython zerver/migrations/02* will find many good examples. Before writing migrations of this form, you should read Django’s docs and the sections below.

  • 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 rebase to do this in the commit that changed models/*.py if 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 add commands, etc.).

  • Release branches: When a release branch needs a migration, but main already has new migrations, the migration graph must fork. The migration should be named and numbered in main to follow in usual sequence, but its dependency should be set to the last migration that exists on the release branch. This should be followed, on main, by a migration which merges the two resulting tips; you can make such a merge with manage.py makemigrations --merge.

  • Large tables: For our very largest tables (e.g., Message and UserMessage), we often need to take precautions when adding columns to the table, performing data backfills, or building indexes. We have a zerver/lib/migrate.py library to help with adding columns and backfilling data.

  • Adding indexes. Django’s regular AddIndex operation (corresponding to CREATE INDEX in SQL) locks writes to the affected table. This can be problematic when dealing with larger tables in particular and we’ve generally preferred to use AddIndexConcurrently (corresponding to CREATE INDEX CONCURRENTLY) to allow the index to be built while the server is active.

  • 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, for example, 1000 objects (e.g., a Message or UserMessage table change). There is a useful Django feature that makes it possible to add atomic=False at the top of a Migration class 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.

  • No-op migrations. Django detects model changes that does not necessarily lead to a schema change in the database. For example, field validators are a part of the Django ORM, but they are not stored in the database. When removing such validators from an existing model, nothing gets dropped from the database, but Django would still generate a migration for that. We prefer to avoid adding this kind of no-op migrations. Instead of generating a new migration, you’ll want to modify the latest migration affecting the field.

  • 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 zerver or 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.models for the reasons above. In this situation you should use Django’s apps.get_model to 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 properties like Realm.subdomain or anything not related to the Django ORM.

    Another important note is that making changes to the data in a table via RunPython code and ALTER TABLE operations within a single, atomic migration don’t mix well. If you encounter an error such as

    django.db.utils.OperationalError: cannot ALTER TABLE "table_name" because it has pending trigger events
    

    when testing the migration, the reason is often that these operations were incorrectly mixed. To resolve this, consider making the migration non-atomic, splitting it into two migration files (recommended), or replacing the RunPython logic with pure SQL (though this can generally be difficult).

  • 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 several commits that are each individually correct, and can each be deployed independently:

      1. 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)

      2. Second, do a migration to copy values from the old column to the new column, to ensure that the two data stores agree.

      3. Third, a commit that stops writing to the old field.

      4. 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 the subject.

We have integrated this system with our test framework so that if you use the use_db_models decorator, you can use some helper methods from 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.

Schema and initial data changes

If you follow the processes described above, tools/provision and tools/test-backend should detect any changes to the declared migrations and run migrations on (./manage.py migrate) or rebuild the relevant database automatically as appropriate.

Notably, both manage.py migrate (git grep post_migrate.connect for details) and restarting tools/run-dev will flush memcached, so you shouldn’t have to worry about cached objects from a previous database schema.

While developing migrations, you may accidentally corrupt your databases while debugging your new code. You can always rebuild these databases from scratch:

  • Use tools/rebuild-test-database to rebuild the database used for test-backend and other automated tests.

  • Use tools/rebuild-dev-database to rebuild the database used in manual testing.