Horton

Re: pgstat_send_connstats() introduces unnecessary timestamp and UDP overhead

Laurenz Albe <laurenz.albe@cybertec.at>

View thread View on postgr.es
From: Laurenz Albe <laurenz.albe@cybertec.at>
To: Andres Freund <andres@anarazel.de>, magnus@hagander.net
Cc: Michael Paquier <michael@paquier.xyz>, Kyotaro Horiguchi <horikyota.ntt@gmail.com>, pgsql-hackers@postgresql.org
Date: 2021-09-08T04:11:56Z
Lists: pgsql-hackers

Attachments

On Fri, 2021-09-03 at 17:04 -0700, Andres Freund wrote:
> Here's how I think that would look like. While writing up this draft, I found
> two more issues:
> 
> - On windows / 32 bit systems, the session time would overflow if idle for
>   longer than ~4300s. long is only 32 bit. Easy to fix obviously.
> - Right now walsenders, including database connected walsenders, are not
>   reported in connection stats. That doesn't seem quite right to me.
> 
> In the patch I made the message for connecting an explicitly reported message,
> that seems cleaner, because it then happens at a clearly defined point. I
> didn't do the same for disconnecting, but perhaps that would be better? Then
> we could get rid of the whole pgStatSessionEndCause variable.

I have gone over your patch and made the following changes:

- cache the last report time in a static variable pgLastSessionReportTime
- add a comment to explain why we only track normal backends
- pgindent
- an attempt at a commit message

Yours,
Laurenz Albe

Commits

  1. Fix performance regression from session statistics.

  2. Add pg_stat_database counters for sessions and session time