The change partly goes back to the old behavior for forwards migrations
which should reduce the amount of memory consumption (#24745). However,
by the way the current state computation is done (there is no
`state_backwards` on a migration class) this change cannot be applied to
backwards migrations. Hence rolling back migrations still requires the
precomputation and storage of the intermediate migration states.
This improvement also implies that Django does not handle mixed
migration plans anymore. Mixed plans consist of a list of migrations
where some are being applied and others are being unapplied.
Thanks Andrew Godwin, Josh Smeaton and Tim Graham for the review as well
as everybody involved on the ticket that kept me looking into the issue.
The new attribute is checked when the `migrate --fake-initial` option
is used. initial will be set to True for all initial migrations (this
is particularly useful when initial migrations are split) as well as
for squashed migrations.
If the only manager on the model is the default manager defined
by Django (`objects = models.Manager()`), this manager will not
be added to the model state. If it is custom, it needs to be
passed to the model state.
The idea behind this change is, that AlterUniqueTogether,
AlterIndexTogether and AlterOrderWithRespectTo can always be moved after
an Add/Alter/Rename/RemoveField operation if they don't refer to the
respective field and are not empty sets / None.
Combined with the optimizations of duplicate AlterUniqueTogether,
AlterIndexTogether, and AlterOrderWithRespectTo operations from
128caa1e16, these operations are optimized
in a later round of the optimizer.
Thanks Tim Graham for the review.
This is a regression caused by introducing rendered migration states in
1aa3e09c20 and the _meta refactoring in fb48eb0581.
Thanks to Danilo Bargen for reporting the issue and Marten Kenbeek and
Tim Graham for triaging the bug and providing the initial test case.
Currently some of the migrations tests rely on the fact 'input' is aliased
because of six, instead of using mock.patch. Replace this code with proper
use of mock.patch.
Also, replace one case of excessively specific python version check with
testing six.PY3
This also prevents state modifications from corrupting previous states.
Previously, when a model defining a relation was unregistered first,
clearing the cache would cause its related models' _meta to be cleared
and would result in the old models losing track of their relations.
