Transaction Manager Brokenness

[Toby Dickenson]

From the "Transaction Manager Changes" thread on zodb-dev.

>TD> I am seeing some unusual behaviour in some classes which hook
>TD> into the transaction manager since Jeremy's sortKey changes. I
>TD> have not been able to reproduce a minimal test case yet, so this
>TD> has been diagnosed from activity logs on a production system.
>
>TD> The following two call sequences are ringing my alarm bells:
>
>TD> 1. abort, abort (that is, abort called twice on
>TD> the same _p_jar object with no intermediate calls)
>
>TD> 2. tpc_begin, abort, tpc_abort
>
>TD> Does this indicate a bug somewhere in the transaction manager,
>TD> or should my objects be expecting this?
>
>TD> (if it helps, there may be subtransactions involved too)
>

[Jeremy Hylton (jeremy-alum)]

Status: Pending => Accepted

 Supporters added: jeremy

It might help to know more about what the resource manager (aka jar) is actually doing. For example, abort() is called with two arguments -- an object and a transaction. What objects are getting passed to abort? Is it the same object twice or is it two different objects?

[Jeremy Hylton (jeremy-alum)]

The tpc_begin(), abort(), tpc_abort() sequence looks normal to me. The transaction got started, something went wrong in the middle (like a ConflictError), so the current transaction was aborted. An abort during 2PC requires that we call abort() for each object and then call tpc_abort().

Is there some specific problem or unusual behavior you are seeing? Or are you just nervous about log messages you haven't seen before?

[Toby Dickenson (htrd)]

> Is there some specific problem or unusual behavior you are seeing? Or
> are you just nervous about log messages you haven't seen before?

I am fairly sure that my objects have never experienced this call sequence before this recent change. Does that qualify as 'unusal behaviour'?

My objects are not seeing any *bad* behaviour, but only thanks to defensive programming. My log messages come from branches that I never expected to be taken.

> The tpc_begin(), abort(), tpc_abort() sequence looks normal to

Really? Having the abort() in the middle *looks* wrong to me, in the same way that mismatched parentheses *look* wrong.

(As I said before, I am not entirely confident that my expectations are correct here)

[Florent Guillaume (efge)]

This will be closed unless details are added, because this is very old and vague.
Also ZODB has changed a lots since then.

[Florent Guillaume (efge)]

Changes: submitter email, edited transcript, importance (critical => low)