Opened 4 years ago

Closed 18 months ago

Last modified 13 months ago

#16969 closed Cleanup/optimization (fixed)

Avoid selecting production database prior to test database setup/teardown

Reported by: andreas_pelme Owned by: nobody
Component: Testing framework Version: master
Severity: Normal Keywords:
Cc: Triage Stage: Ready for checkin
Has patch: yes Needs documentation: no
Needs tests: no Patch needs improvement: no
Easy pickings: no UI/UX: no


I am playing around with running my test suite with pytest and pytest-xdist. I would like to be able to run my tests in parallell, which would probably be a massive speedup on a decent multi-core CPU. To isolate tests properly, I am dynamically modifying the database name in the settings, to create a separate test database for each process.

However, when the test database is created, Django creates a cursor to the production database before setting up/destroying the test database. Since I am generating the database name dynamically in each process, this causes problems for me, since I do not know the name and can create it beforehand. I see no reason for doing this, so I would like to skip selecting the production database before creating/destroying the test database.

The behavior is when creating a cursor without specifying a database varies a bit across the backends:

  • sqlite - this is not a problem
  • mysql - it is possible to omit the database name
  • postgesql - a database name must be given, but rather than choosing the production name, the default "postgesql" database can be used
  • oracle - I am unable to test this since I do not have an Oracle database available

I have modified django/db/backends/ to set connection.settings_dictNAME? to None instead of the production database name during creation/destroy. I also made a small change in the postgres backend to use the "postgesql" database if None is supplied.

I have run the Django test suite for sqlite, mysql and postgesql and this does not seems to cause any problems.

Attachments (2)

test_db_creation.diff (2.2 KB) - added by andreas_pelme 4 years ago.
16969-2.diff (4.4 KB) - added by claudep 19 months ago.

Download all attachments as: .zip

Change History (15)

Changed 4 years ago by andreas_pelme

comment:1 Changed 4 years ago by carljm

  • Has patch set
  • Needs documentation unset
  • Needs tests unset
  • Patch needs improvement unset
  • Triage Stage changed from Unreviewed to Accepted

If its possible to avoid connecting to the production DB at all for test runs, I think we should do this. Needs someone with Oracle to test the behavior there.

comment:2 Changed 4 years ago by aaugustin

The test suite runs under Oracle with the patch applied.

comment:3 Changed 4 years ago by aaugustin

Under PostgreSQL, is it guaranteed that all users can always connect to the postgres database? Otherwise, the patch is incorrect.

comment:4 Changed 19 months ago by timo

Regarding PostgreSQL, the answer appears to be "Yes". From the PostgreSQL docs:

"The postgres database is a default database meant for use by users, utilities and third party applications."

Additional information in this stackoverflow question.

Changed 19 months ago by claudep

comment:5 Changed 19 months ago by claudep

Here's an alternative patch. Should be thoroughly reviewed!

comment:6 Changed 19 months ago by shai

Oracle does not support "databases" in the sense that other backends do; the NAME parameter is actually part of selecting the server you connect to. The tests worked with the patch because the patch changed the generic create_test_db function, which the Oracle backend overrides. In fact, Oracle doesn't use the TEST_NAME at all (except for printing) -- it uses a TEST_USER and TEST_TBLSPACE instead.

To summarize: On Oracle, you connect as a user to a server. There is only one "database". You cannot avoid connecting to it, AFAIK.

comment:7 Changed 19 months ago by claudep

In my patch, then, what about subclassing nodb_connection() in the Oracle backend and returning self.connection?

comment:8 Changed 19 months ago by shai

There's no need for that, the Oracle backend already overrides _create_test_db() as well as _destroy_test_db(), so it wouldn't use nodb_connection().

On point that rises, though, is that nodb_connection() should probably be named as a private method (_nodb_connection()).

comment:9 Changed 19 months ago by claudep

Pull request created (using _nodb_connection() private name):

comment:10 Changed 18 months ago by timo

  • Triage Stage changed from Accepted to Ready for checkin

comment:11 Changed 18 months ago by Claude Paroz <claude@…>

  • Resolution set to fixed
  • Status changed from new to closed

In e953c78eeb81ee69dccd356145563fd6f9e4c7b6:

Fixed #16969 -- Don't connect to named database when possible

Thanks Andreas Pelme for the report and initial patch, and
Aymeric Augustin, Shai Berger and Tim Graham for the reviews.

comment:12 Changed 18 months ago by Claude Paroz <claude@…>

In 7e714827ead50f77aa82394cc2511ff96ab67fa4:

Don't initialize PostGIS-specific stuff for non-db connections

Refs #16969.

comment:13 Changed 13 months ago by Tim Graham <timograham@…>

In 0086c9eb48101b6dabefa7f9195ac7d197780f09:

[1.7.x] Fixed #22417 -- Added additional documentation for refs #16969.

Thanks Jon Foster for the report.

Backport of 1b3a3fc1e4 from master

Note: See TracTickets for help on using tickets.
Back to Top