Re: idle_in_transaction_timeout

On Thu, Jun 26, 2014 at 12:40 AM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> Fujii Masao <masao(dot)fujii(at)gmail(dot)com> writes:
>> Why is IIT timeout turned on only when send_ready_for_query is true?
>> I was thinking it should be turned on every time a message is received.
>> Imagine the case where the session is in idle-in-transaction state and
>> a client gets stuck after sending Parse message and before sending Bind
>> message. This case would also cause long transaction problem and should
>> be resolved by IIT timeout. But IIT timeout that this patch adds cannot
>> handle this case because it's enabled only when send_ready_for_query is
>> true. Thought?
>
> I think you just moved the goalposts to the next county.
>
> The point of this feature, AFAICS, is to detect clients that are failing
> to issue another query or close their transaction as a result of bad
> client logic. It's not to protect against network glitches.

If so, the document should explain that cleanly. Otherwise users may
misunderstand this parameter and try to use it to protect even long transaction
generated by network glitches or client freeze.

> Moreover, there would be no way to implement a timeout like that without
> adding a gettimeofday() call after every packet receipt, which is overhead
> we do not need and cannot afford.

Hmm.. right. I was thinking to just call enable_timeout_after() every time
message is received. But it calls gettimeofday() and causes overhead.

> I don't think this feature should add
> *any* gettimeofday's beyond the ones that are already there.

But, ISTM that the patch adds enable_timeout_after() which calls
GetCurrentTimestamp()->gettimeofday() before receiving new message
when send_ready_for_query is true. When send_ready_for_query is true,
it's expected that next message takes time to arrive at server, so
calling gettimeofday() in that case might not be so harmful, though.

Regards,

--
Fujii Masao