Re: [HACKERS] WIP Patch: Pgbench Serialization and deadlock errors

On Thu, Jun 15, 2017 at 2:16 PM, Andres Freund <andres@anarazel.de> wrote:
> On 2017-06-14 11:48:25 +0300, Marina Polyakova wrote:

>> I suggest a patch where pgbench client sessions are not disconnected because
>> of serialization or deadlock failures and these failures are mentioned in
>> reports.
>
> I think that's a good idea and sorely needed.

+1

>> P.S. Does this use case (do not retry transaction with serialization or
>> deadlock failure) is most interesting or failed transactions should be
>> retried (and how much times if there seems to be no hope of success...)?
>
> I can't quite parse that sentence, could you restate?

The way I read it was that the most interesting solution would retry
a transaction from the beginning on a serialization failure or
deadlock failure.  Most people who use serializable transactions (at
least in my experience) run though a framework that does that
automatically, regardless of what client code initiated the
transaction.  These retries are generally hidden from the client
code -- it just looks like the transaction took a bit longer.
Sometimes people will have a limit on the number of retries.  I
never used such a limit and never had a problem, because our
implementation of serializable transactions will not throw a
serialization failure error until one of the transactions involved
in causing it has successfully committed -- meaning that the retry
can only hit this again on a *new* set of transactions.

Essentially, the transaction should only count toward the TPS rate
when it eventually completes without a serialization failure.

Marina, did I understand you correctly?

-- 
Kevin Grittner
VMware vCenter Server
https://www.vmware.com/