[ TL;DR: A migration may use a “replaces” list pointing to migrations that don’t actually exist. This undocumented technique cleanly solves a recurring difficult migration problem. We seek consensus on whether this should become a documented feature or that it is an unexpected side effect subject to change in the future. ]
We have found an undocumented behavior in the migration system that gracefully solves the troublesome problem of merging migrations created in parallel development branches. If this behavior should survive, we’ll enter a documentation ticket – but if it’s considered a bug, we’ll need to stay away from it and fall back to the more difficult manual editing approaches we’ve used in the past. The Use Case ------------------ We’re rapidly developing a large multi-tenant application (hundreds of ORM models, thousands of migrations and hundreds of thousands of lines of code so far, with quite a bit of work remaining) punctuated by periodic production releases. We create a source code branch from our mainline development trunk for each production release, just in case we must rapidly issue patches to those production releases. On rare occasions, we’ve had to make a schema change (such as adding a new field) as a patch to a production release, and make a parallel schema change in the mainline development trunk. Of course, this normally causes a migration failure when migrating a production tenant from the patch release up to a later version of the mainline release – since the mainline release has a subsequent migration that adds the same field. We’ve solved this in the past by manually rearranging the dependency order of the mainline trunk migrations (moving the replacement step before other new migrations for this later release), and fiddling with the contents of the django_migrations table to make it look like that mainline step has already been run before running the migrations. We’re unhappy with that approach – it’s both time consuming and error prone. This problem is similar to, but not identical to, that of squashing migrations. (And yes, we do periodically squash our migrations. We have about 600 migration steps at the moment, left over from more than 2,000 originally created. We’ve got another round of squashing coming up soon that should take us to less than 100 migrations – but we have more than a dozen developers adding more migrations every week.) The Discovery ------------------- Through trial and error, we found that our mainline migration step may declare itself as a replacement for the patch step (using the “replaces” attribute) – even if the patch migration itself doesn’t exist in the list of mainline migrations. And if we do this, the migration engine simply works as hoped and our problem vanishes. It’s absolutely wonderful; simple to implement and effective. We love it. New tenants run only the replacement step; tenants migrating from the patch release to the trunk release merely record the replacement step as having been completed without actually executing it; development tenants that never saw the original patch step simply record both the patch step and the replacement as having been completed. It’s great. The Worry -------------- This approach seems undocumented in three different ways: * The replacement migration is pointing at an original migration that doesn’t exist in the trunk’s migration files. (We created it in the patch branch and we know the migration name from that branch, but we never added the patch migration to the mainline trunk.) The current documentation[1] describes keeping both the original and the replacement in place until all databases have migrated past the replacement step (and then deleting the original and removing the “replaces” attribute from the replacement). The documentation implies, but does not explicitly state, that the original step should exist in the list. Our testing shows that the original need not exist (and we like it this way!). * If we go ahead and add a copy of the patch release’s migration step to the mainline trunk, we introduce a “multiple leaf nodes” graph, since none of the mainline migrations depend upon this “side patch”. However, apparently because there is a declared replacement for this patch step, the migration engine doesn’t raise the “multiple leaf nodes” exception. This seems to be an oversight unless the replacement step is somehow acting as a merge (as if it had a dependency on the patch step) … but we like the way it’s working now, if it were to become necessary to include the original step in the mainline migration list. * We have found that we can have multiple replacement steps all claiming to replace the same original step number. (This conveniently handled a case where multiple migrations were originally created in the trunk, then backported as a single migration into a patch to an earlier production release.) But this results in the path migration’s app and name being inserted into django_migrations table more than once. These duplicate entries haven’t appeared to cause a problem, but they were unexpected. It seems that the app and migration name ought to be “unique together” but aren’t – perhaps for performance reasons, since the contents of this table are normally managed solely by the migrations system. The Question ------------------- Would the core team consider the ability to “replace” a non-existent migration step to be a feature or a bug? We prefer to think of this as a desirable feature, since it solves what seems to be a non-uncommon use case. We haven’t seen any other documented approaches to solving the problem of migrations created in parallel branches – most published advice boils down to either “don’t do it”, “roll back your migrations then apply the new ones”, or “good luck on manually repairing things.” If this IS considered a bug, we certainly could add the original migration from the patch release, but then we’ve added a migration “to the side” of the original dependency tree introducing another leaf node. We’d hate for that to be considered a problem in the future, because the replacement step doesn’t look like it should act as a merge node (it doesn’t depend upon the original, just replaces it). The third point, the insertion of duplicate records into django_migrations, does smell like a defect. If people like this “feature” and believe it should be supported, we’d be happy to create a documentation PR. Barry Johnson Epicor Software Corporation [1]: https://docs.djangoproject.com/en/2.2/topics/migrations/#migration-squashing -- You received this message because you are subscribed to the Google Groups "Django developers (Contributions to Django itself)" group. To unsubscribe from this group and stop receiving emails from it, send an email to django-developers+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/django-developers/29136C68-DA75-431E-8C77-169097346AD1%40epicor.com.