Squash migrations in django.contrib.auth for better performance and nicer output

[Tim Abbott]

Provisioning a fresh database for any project using Django's contrib.auth system ends up spending about 0.3 seconds, and probably more importantly a nontrivial amount of output on the screen running the 12 migrations in django.contrib.auth in sequence. This can be optimized significantly by squashing the migrations using the manage.py squashmigrations algorithm.

Sample output using ts "%b %d %H:%M:%.S" for timing output:

Aug 23 17:07:23.116481 Applying auth.0002_alter_permission_name_max_length... OK
Aug 23 17:07:23.131759 Applying auth.0003_alter_user_email_max_length... OK
Aug 23 17:07:23.149192 Applying auth.0004_alter_user_username_opts... OK
Aug 23 17:07:23.166695 Applying auth.0005_alter_user_last_login_null... OK
Aug 23 17:07:23.169690 Applying auth.0006_require_contenttypes_0002... OK
Aug 23 17:07:23.191648 Applying auth.0007_alter_validators_add_error_messages... OK
Aug 23 17:07:23.207522 Applying auth.0008_alter_user_username_max_length... OK
Aug 23 17:07:23.234117 Applying auth.0009_alter_user_last_name_max_length... OK
Aug 23 17:07:23.332708 Applying auth.0010_alter_group_name_max_length... OK
Aug 23 17:07:23.417119 Applying auth.0011_update_proxy_permissions... OK
Aug 23 17:07:23.433407 Applying auth.0012_alter_user_first_name_max_length... OK

The django.contrib.auth.migrations tree hasn't see any new migrations since migration 0012_alter_user_first_name_max_length.py was added in 2020, so it seems fairly likely that this won't need to be done on a recurring basis.

[Jacob Walls]

Saving 0.3s from test runs is great, but it's worth noting this will impose a breaking change on users who have explicit dependencies on these migrations in their own migration files.

[Jacob Walls]

Here is a bench from my development machine.

Python 3.12.1
Apple M2 Pro
Django 5.2.a-0@cdcd604ef8f650533eff6bd63a517ebb4ffddf96
Postgres 15.4

Current: 0.137s

Running migrations:
  Applying auth.0001_initial... OK (0.071s)
  Applying auth.0002_alter_permission_name_max_length... OK (0.005s)
  Applying auth.0003_alter_user_email_max_length... OK (0.003s)
  Applying auth.0004_alter_user_username_opts... OK (0.003s)
  Applying auth.0005_alter_user_last_login_null... OK (0.004s)
  Applying auth.0006_require_contenttypes_0002... OK (0.001s)
  Applying auth.0007_alter_validators_add_error_messages... OK (0.003s)
  Applying auth.0008_alter_user_username_max_length... OK (0.010s)
  Applying auth.0009_alter_user_last_name_max_length... OK (0.003s)
  Applying auth.0010_alter_group_name_max_length... OK (0.004s)
  Applying auth.0011_update_proxy_permissions... OK (0.026s)
  Applying auth.0012_alter_user_first_name_max_length... OK (0.004s)

Squashed: 0.093s

Running migrations:
  Applying auth.0001_initial_squashed_0012_alter_user_first_name_max_length... OK (0.093s)

---

That's a savings of about 1/3 of the migration time, or 0.044s on an admittedly fast machine (so double it for cloud CI). Given that this would impose a breaking change on users (and the verbosity can already be silenced with -v0), I don't see the value outweighing the cost. We also have guidance in the migrations docs about not worrying about the number of migrations until you have ​hundreds.

[Natalia Bidart]

Hello Tim, thank you for your ticket! I agree with the resolution from this ticket and with the commentary from Jacob.

Squashing auth migrations without a proper deprecation path, and without considering all the possible scenarios of usage by Django projects, will likely be a breaking change and the gain does not feel worth producing that breakage.

Because of the above, this report should be treated similarly to a "new feature request" for Django. For cases like this, the recommended path forward is to first propose and discuss the idea/request with the community and gain consensus. To do that, please consider starting a new conversation on the ​Django Forum, where you'll reach a wider audience and likely get extra feedback.

If the community agrees with the proposal, please return to this ticket and reference the forum discussion so we can re-open it. For more information, please refer to ​the documented guidelines for requesting features.

[Tim Abbott]

Thanks for investigating!

this will impose a breaking change on users who have explicit dependencies on these migrations in their own migration files

Hmm, I wasn't aware of any breaking change concern with squashing migrations, as long as one doesn't delete the original migrations when doing the squashing. I'll see if I can find time to start a proper Django forum discussion.

[Jacob Walls]

as long as one doesn't delete the original migrations when doing the squashing.

Sorry, I took that to be part of the point of the proposal, but it looks like that was an assumption on my part.

[Natalia Bidart]

Replying to Jacob Walls:

as long as one doesn't delete the original migrations when doing the squashing.

Sorry, I took that to be part of the point of the proposal, but it looks like that was an assumption on my part.

I did assume that the original migrations would not be deleted. But, my thinking when mentioning breakages was more related to: when new auth migrations need to be added, which migration would these new files depend on? the latest non-squashed? the squashed one? As far as I understand either answer breaks "the other case" (our docs are not that explicit about a case similar to this).

Because of the impact that this can (potentially) have, I think that, if we want to consider re-opening, we need a forum discussion with some general agreement that this is both:

• desired, and
• does not break current or future Django projects.