[Sqlalchemy-tickets] Issue #2964: Race condition in dialect initialization (zzzeek/sqlalchemy)

[dairiki <iss...@bi...>]

New issue 2964: Race condition in dialect initialization
https://bitbucket.org/zzzeek/sqlalchemy/issue/2964/race-condition-in-dialect-initialization

dairiki:

There is a race condition  in `DefaultEngineStrategy` around [line 161](https://bitbucket.org/zzzeek/sqlalchemy/src/fdedb69d9c6d995fc85105c22f0f537ea25a0c44/lib/sqlalchemy/engine/strategies.py?at=master#cl-161)
where `dialect.initialize()` gets called from a listener attached to the pool `first_connect` event.

If there are multiple threads all rushing to query the database, however,
the second thread to get a connection can start running queries before the first thread has finished initializing the dialect.

(In my case, `dialect.returns_unicode_strings` was intermittently not being initialized before the first queries were performed.  This resulted in a UnicodeResultProcessor being used when it shouldn't have been in turn leading to UnicodeEncodeError.)

Not sure what the best fix is.  My guess is that `dialect.initialized` should be hooked into the `connect` event instead of `first_connect`, with a wrapper which ensures that `dialect.initialize` only gets called once by the first thread to try — blocking subsequent threads until the initialization is complete.  (The locking/blocking could perhaps be folded into `util.only_once`.)