Documentation: db transactions clarification

[startup.canada@…]

In section ​https://docs.djangoproject.com/en/dev/topics/db/transactions/#requirements-for-transaction-handling says "Django requires that every transaction that is opened is closed before the completion of a request. If you are using autocommit() (the default commit mode) or commit_on_success(), this will be done for you automatically.", giving the impression that given the autocommit we never need to perform commits, but this is not correct for the case of writing raw sql inserts, as explained in ​https://docs.djangoproject.com/en/dev/topics/db/sql/#executing-custom-sql-directly "After performing a data changing operation, you should then call transaction.commit_unless_managed() to ensure your changes are committed to the database".

Just for clarification I suggest adding in the #requirements-for-transaction-handling a sentence mentioning the need for manually committing for raw sql inserts and/or linking to the executing-custom-sql-directly section.

[Tim Graham <timograham@…>]

In 31f49f1396c4cc565017066e9f17c6b850a89687:

Fixed #19442 - Clarified that raw SQL must be committed.

Thanks startup.canada for the suggestion.

[Tim Graham <timograham@…>]

In e11bd5c5ac0c7d85723451f8da7b6ad30fefc680:

[1.5.X] Fixed #19442 - Clarified that raw SQL must be committed.

Thanks startup.canada for the suggestion.

Backport of 31f49f1396 from master

[Anssi Kääriäinen]

The modified docs suggest that both autocommit and commit_on_success need manual commit for raw SQL. I don't think this is the case - if you are using raw SQL, then commit_on_success should just work, but autocommit fails to commit your raw SQL writes (as it doesn't know which raw SQL statements are writes...).

[Tim Graham]

I guess the linked section (/topics/db/sql/#executing-custom-sql-directly) should be clarified as well as it seems to suggest you need to call transaction.commit_unless_managed regardless of transaction mode.

[Anssi Kääriäinen]

It is possible that I just remember this wrong, though I am pretty sure the behaviour is as I said in comment:4. I will try to double-check how the decorators work later today. It is also possible that the behaviour has changed along the years and not all parts of the docs reflect this.

[Anssi Kääriäinen]

The added test case (not meant to be included in the test suite) shows how the autocommit and commit_on_success work. Basically, autocommit is somewhat dangerous, and somewhat easily leads to transaction leak.

Anyhow, the point for this ticket is that no, you don't need to commit raw SQL when using commit_on_success.

I do think it would be a good idea to rewrite Django's transaction handling. It currently works nicely in request processing, somewhat works in single db shell scripts and is really hard to use in multidb shell scripts. But this of course isn't this ticket's problem.

[Tim Graham]

Thanks for your research! Updated patch attached. Seems that Aymeric doesn't like the idea of rewriting transaction management for offline tasks (see #13870). I'd be appreciative if you'd like to review the documentation patch on that ticket as well.

[Anssi Kääriäinen]

Maybe rewrite of transaction management is a little drastic, but adding some more helpers at least would make sense.

We would likely want to add docs of how to use transactions safely in Celery jobs (or other long running processes which aren't serving requests). Here the basic idea is that if you make sure _all_ connections are closed after a job is ran, then you can proceed like in request processing. If you don't close all connections, then you must be extra careful in making sure no transactions are left open. #13870 is of course the right ticket for work on this area.

I will take a closer look at the whole transaction management documentation. The patch is definitely an improvement to current docs, but I wonder if there is something more lurking in there...

[Tim Graham]

Shall we commit this patch, mark it as needs improvement, and/or open a ticket for following up with additional doc work per your comment?

[Tim Graham]

Don't think this is relevant anymore with the merge of the database-level-autocommit branch [14cddf51c5].