Re: [INTERFACES] Transaction support in 6.5.3/JDBC

Peter Mount <petermount(at)it(dot)maidstone(dot)gov(dot)uk> writes:
> 1. When a dead-lock occurs trying to update the same row from multiple
> threads, there does not seem to be any timeout and both connections hang
> forever. Is there any dead-lock detection in 6.5 and will there be on in
> 6.6 or 7.0?

> PM: I'm not sure if there is/will be in the backend. I'm planning on
> making the OOB abort stuff working for 7.0, but I'm not certain on
> dead-lock's.

This doesn't seem right to me --- the backend should detect deadlocks
(not right away, but within a few seconds or a minute at most). If
it doesn't, that's not JDBC's fault. More details please?

> 2. In the event of such a dead-lock, the XA layer will attempt to
> terminate one of the connections. Right now the only recourse is to
> shutdown the connection forcefully, hoping that the transaction is
> rolledback and all locks are released. Am I safe to assume that?

> PM: Someone correct me if I'm wrong, but the transaction should roll
> back if the connection is closed whilst it's open, regardless of the
> interface.

Right. Again, this'd be a backend bug if it didn't happen. The backend
isn't supposed to rely on correct frontend behavior to ensure database
integrity.

> 3. I would rather cancel the pending update/insert, and according to the
> interface specs there is a way to recieve a pid/key on startup and use
> it to cancel an operation in progress. However, at the end of the
> authentication process that happens at startup, no pid/key are send to
> the FE, nor does the BE acknowledge that a query can be made.

> PM: Do you mean the pid of the backend running a particular connection?
> If so, I can see some security headaches if one connection can cancel an
> operation running on another. As for the BE not acknowledging that a
> query can be made, thats just how the current protocol works.

This one maybe can be blamed on JDBC. In the 2.0 protocol the BE
will send a cancel security key --- but I seem to recall that the
JDBC client still requests protocol 1.0?

regards, tom lane