Re: Fix around conn_duration in pgbench - Mailing list pgsql-hackers

On Fri, 20 Aug 2021 02:05:27 +0900
Fujii Masao <masao.fujii@oss.nttdata.com> wrote:

> 
> On 2021/08/11 13:56, Fujii Masao wrote:
> > Yes, but I was thinking that's a bug that we should fix.
> > IOW, I was thinking that, in v13, both connection and disconnection delays
> > should be measured whether -C is specified or not, *per spec*.
> > But, in v13, the disconnection delays are not measured in the cases
> > where -C is specified and not specified. So I was thinking that this is
> > a bug and we should fix those both cases.
> > 
> > But you're thinking that, in v13, the disconnection delays don't need to
> > be measured because they are not measured for now?
> 
> Please let me clarify my thought.

Thank you for your clarification.

> 
> In master and v14,
> 
> # Expected behavior
> (1) Both connection and disconnection delays should be measured
>        only when -C is specified, but not otherwise.
> (2) When -C is specified, since each transaction establishes and closes
>        a connection, those delays should be measured for each transaction.
> 
> # Current behavior
> (1) Connection delay is measured whether -C is specified or not.
> (2) Even when -C is specified, disconnection delay is NOT measured
>        at the end of transaction.
> 
> # What the patch should do
> (1) Make pgbench skip measuring connection and disconnection delays
>        if not necessary (i.e., -C is not specified).
> (2) Make pgbench measure the disconnection delays whenever
>        the connection is closed at the end of transaction, when -C is specified.

I agree with you. This is what the patch for pg14 does. We don't need to measure
disconnection delay when -C is not specified because the output just reports
"initial connection time".

> In v13 or before,
> 
> # Expected behavior
> (1) Both connection and disconnection delays should be measured
>        whether -C is specified or not. Because information about those delays
>        is used for the benchmark result report.
> (2) When -C is specified, since each transaction establishes and closes
>        a connection, those delays should be measured for each transaction.
> 
> # Current behavior
> (1)(2) Disconnection delay is NOT measured whether -C is specified or not.
> 
> # What the patch should do
> (1)(2) Make pgbench measure the disconnection delays whenever
>              the connection is closed at the end of transaction (for -C case)
>              and the end of thread (for NOT -C case).

Ok. That makes sense. The output reports "including connections establishing"
and "excluding connections establishing" regardless with -C, so we should
measure delays in the same way.

I updated the patch for pg13 to measure disconnection delay when -C is not
specified. I attached the updated patch for pg13 as well as one for pg14
which is same as attached before.

> 
> Anyway, I changed the status of this patch to "Waiting on Author" in CF.

I returned the status to "Ready for Committer". 
Could you please review this?

Regards,
Yugo Nagata

-- 
Yugo NAGATA <nagata@sraoss.co.jp>