gevent.local.local breaks transaction.manager about transaction

[jamadden]

Hi! I'm a maintainer of gevent and I've also contributed to this library. My organization runs a gevent-patched system that heavily uses transaction and I don't recall that we've run into any trouble. I also don't think we have any special code, aside from an assert in one of our modules:

assert issubclass(transaction.ThreadTransactionManager, gevent.local.local)

This makes sure that the monkey-patching has been done soon enough. The trick, as gevent tries to document, is to make sure that monkey-patching is done immediately during process startup, before any other imports. In our case, we have modified pserve to do the patching, but that might differ in a given environment.

Our assert code is slightly more general than that to account for other thread-local things we're aware of. I can share it if that would be helpful.

[miohtama]

I moved gevent.monkey.patch_all() to very early in my program and managed to resolve the issue.

[miohtama]

@jamadden Here is the description of the sympton: ThreadTransactionManager gives same _txn even if the thread differs.

Thread is <ServiceThread(xxx, started 4449301920)>, manager is <transaction._manager.ThreadTransactionManager object at 0x1072e0e88>, Tx is <transaction._transaction.Transaction object at 0x109373fd0>
Thread is <ServiceThread(yyy, started 4449302072)>, manager is <transaction._manager.ThreadTransactionManager object at 0x1072e0e88>, Tx is <transaction._transaction.Transaction object at 0x109373fd0>

This is using threads started with threading.Thread.start.

I'd assume I need to get patched threading.Thread?