Safety when we get an error during transaction begin

[jzuech3]

Resolves #18

[jamadden]

Conflicts with the more extensive changes in #21, which should also fix #18. (The cause of #18 was the del tx, not it being 'None'.)

[jzuech3]

Agreed. Merge that first. Even if we remove the del tx, we'll need to handle tx being None in _retryable.

[jamadden]

Nah, if txm.begin() fails, something terrible happened. It's unlikely that it'd be a retryable exception. That just needs to be outside the scope of the retry logic. (RelStorage even has special logic to make sure that if something retryable does happen during txm.begin() -> Connection.newTransaction -> storage.poll_invalidations() it gets deferred until later; no one expects txm.begin() to fail. Which, now that I think about, may be slightly incomplete in newer versions...)

[jzuech3]

Ah, ok. We can probably delete this PR then. #18 was caused by an issue during tx.begin right?

[jamadden]

#18 was caused by an issue during tx.begin right?

Possibly, but not necessarily. There used to be del tx while we were still inside the scope of the try/except handler, and the arrival of a signal, say, could happen after del tx but before exiting the try/except block. Other conditions detected by the VM like MemoryError could also happen in that window.