Mikko Tiihonen <Mikko.Tiihonen@nitorcreations.com> writes:
> I created a proof of concecpt patch for postgresql JDBC driver that allows the caller to do pipelining of requests
withina transaction. The pipelining here means same as for HTTP: the client can send the next execution already before
waitingfor the response of the previous request to be fully processed.
In principle this can work if you think through the error handling
carefully. The way the FE/BE protocol is intended to handle the case
is that all the pipelined statements are part of one transaction so
that they all succeed or fail together. Otherwise the user has to work
through and predict the outcome of all the combinations of "Q1 succeeded,
Q2 failed, Q3 succeeded" which is mighty error-prone, especially if the
commands are interdependent in any way at all. Having legislated that
(and forbidden any transaction-boundary commands in a pipelined group),
what you do is not issue a Sync until after the last command of the set.
Then, if any command fails, the backend just automatically discards
remaining messages until it gets the Sync, and so you can safely issue
new commands before knowing the results of the previous ones.
I mention this because, although I don't know the innards of the JDBC
driver at all, it sure doesn't look like that's the way you've done it.
If I'm right that you've left the error recovery strategy to the user,
I think this is gonna be very hard to use safely.
regards, tom lane
