Referring to partially run squashed migrations

[Jarek Glowacki]

Consider the migration tree defined by:

my_app/migrations/
0001_initial.py
0002_blah.py -> dep on 0001_initial
0001_initial_squashed_0002_blah.py -> replaces 0001_initial and 0002_blah
0003_ohnoes.py -> dep on ????

If we try to make 0003 depend on 0001_initial_squashed_0002_blah, then we get a NodeNotFoundError if the database is only up to 0001_initial: it is currently in the 'unsquashed reality' in which 0001_initial_squashed_0002_blah does not exist. Creating new databases from this setup does not raise any problems, nor does running the migrations on databases that aren't part-way through a squashed migration series since these will live in the 'squashed reality'.

If we try to make 0003 depend on 0002_blah, then things work fine in all cases.

However makemigrations chooses to point newly created migrations at the squashed migration 0001_initial_squashed_0002_blah rather than the last of the unsquashed ones 0002_blah, causing the problem described above and making it unobvious that new migrations should still point at the old replaced migrations up until the point at which the squashed migrations get transitioned to normal migrations.

Proposed solution:

Fully support listing the dependency both ways, as 0001_initial_squashed_0002_blah and as 0002_blah, using fancy logic to ensure valid references to inapplicable migrations are appropriately rerouted.

​/ Patch
​/ Pull Request

[Tim Graham]

Left comments for improvement on the pull request.

[Markus Holtermann <info@…>]

In 509379a1:

Fixed #25945, #26292 -- Refactored MigrationLoader.build_graph()