django.db.transaction.atomic() leaks savepoints on exception

[hyperair@…]

When savepoints are rolled back after an exception in a transaction.atomic() block, the savepoint is leaked. Specifically, "ROLLBACK TO SAVEPOINT xx" is done, but not "RELEASE SAVEPOINT xx". The savepoint ID is then thrown away by Django and leaked.

See ​http://www.postgresql.org/docs/9.1/static/sql-release-savepoint.html and ​http://www.postgresql.org/docs/9.1/static/sql-rollback-to.html for more information.

When spinning in a get_or_create() loop in a single transaction, this has the potential to max out the shared memory of the PostgreSQL instance resulting in this:

  File "/usr/lib/python2.7/dist-packages/django/db/models/manager.py", line 157, in create
    return self.get_queryset().create(**kwargs)
  File "/usr/lib/python2.7/dist-packages/django/db/models/query.py", line 319, in create
    obj.save(force_insert=True, using=self.db)
  File "/usr/lib/python2.7/dist-packages/django/db/models/base.py", line 545, in save
    force_update=force_update, update_fields=update_fields)
  File "/usr/lib/python2.7/dist-packages/django/db/models/base.py", line 573, in save_base
    updated = self._save_table(raw, cls, force_insert, force_update, using, update_fields)
  File "/usr/lib/python2.7/dist-packages/django/db/models/base.py", line 654, in _save_table
    result = self._do_insert(cls._base_manager, using, fields, update_pk, raw)
  File "/usr/lib/python2.7/dist-packages/django/db/models/base.py", line 687, in _do_insert
    using=using, raw=raw)
  File "/usr/lib/python2.7/dist-packages/django/db/models/manager.py", line 232, in _insert
    return insert_query(self.model, objs, fields, **kwargs)
  File "/usr/lib/python2.7/dist-packages/django/db/models/query.py", line 1511, in insert_query
    return query.get_compiler(using=using).execute_sql(return_id)
  File "/usr/lib/python2.7/dist-packages/django/db/models/sql/compiler.py", line 898, in execute_sql
    cursor.execute(sql, params)
  File "/usr/lib/python2.7/dist-packages/django/db/backends/util.py", line 69, in execute
    return super(CursorDebugWrapper, self).execute(sql, params)
  File "/usr/lib/python2.7/dist-packages/django/db/backends/util.py", line 53, in execute
    return self.cursor.execute(sql, params)
  File "/usr/lib/python2.7/dist-packages/django/db/utils.py", line 99, in __exit__
    six.reraise(dj_exc_type, dj_exc_value, traceback)
  File "/usr/lib/python2.7/dist-packages/django/db/backends/util.py", line 53, in execute
    return self.cursor.execute(sql, params)
django.db.utils.OperationalError: out of shared memory
HINT:  You might need to increase max_locks_per_transaction.

[Chow Loong Jin <hyperair@…>]

Patch to call savepoint_commit after savepoint_rollback in Atomic.exit

[Aymeric Augustin]

The fix was correct. I added a test.

PR: ​https://github.com/django/django/pull/2992

I suspect it will throw some assertNumQueries off. I'll let Jenkins run the tests and see what happens.

I don't plan to backport it to 1.7, because we've already released RC2 and it appears to be an fairly rare case (otherwise it would have been reported earlier). To work around the problem, you can:

• create smaller batches and wrap each batch in its own transaction;
• use a patched copy of Django.

[Aymeric Augustin <aymeric.augustin@…>]

In 729e4ae4f0730585ac4640e7fa3aa06374677ff2:

Fixed #23074 -- Avoided leaking savepoints in atomic.

Thanks Chow Loong Jin for the report and the initial patch.