Horton

Re: Libpq support to connect to standby server as priority

Alvaro Herrera <alvherre@alvh.no-ip.org>

View thread View on postgr.es
From: Alvaro Herrera <alvherre@alvh.no-ip.org>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: Anastasia Lubennikova <a.lubennikova@postgrespro.ru>, Greg Nancarrow <gregn4422@gmail.com>, Peter Smith <smithpb2250@gmail.com>, "Smith, Peter" <peters@fast.au.fujitsu.com>, David Steele <david@pgmasters.net>, "Tsunakawa, Takayuki" <tsunakawa.takay@fujitsu.com>, Haribabu Kommi <kommi.haribabu@gmail.com>, Robert Haas <robertmhaas@gmail.com>, Laurenz Albe <laurenz.albe@cybertec.at>, Dave Cramer <pg@fastcrypt.com>, Jing Wang <jingwangian@gmail.com>, PostgreSQL-development <pgsql-hackers@postgresql.org>, Elvis Pranskevichus <elprans@gmail.com>
Date: 2020-11-24T22:59:17Z
Lists: pgsql-hackers 
On 2020-Nov-24, Tom Lane wrote:

> I'm inclined to go ahead and commit the 0001 patch I posted at [1]
> (ie, change the implementation of GUC_REPORT to avoid intra-query
> reports), since that addresses a performance problem that's
> independent of the goal here.  The rest of this seems to still
> be in Greg's court.

Sounded a good idea to me.

> Has anyone got an opinion about the further improvement I suggested:
> 
> >> As it stands, 0001 reduces the ParameterStatus message traffic to
> >> at most one per GUC per query, but it doesn't attempt to eliminate
> >> duplicate ParameterStatus messages altogether.  We could do that
> >> as a pretty simple adjustment if we're willing to expend the storage
> >> to remember the last value sent to the client.  It might be worth
> >> doing, since for example the function-SET-clause case would typically
> >> lead to no net change in the GUC's value by the end of the query.
> 
> On reflection this seems worth doing, since excess client traffic
> is far from free.

Agreed.  If this is just a few hundred bytes of server-side local memory
per backend, it seems definitely worth it.

Commits

  1. Make test_target_session_attrs more robust against connection failure.

  2. Extend the abilities of libpq's target_session_attrs parameter.

  3. Mark default_transaction_read_only as GUC_REPORT.

  4. Introduce a new GUC_REPORT setting "in_hot_standby".

  5. Avoid spamming the client with multiple ParameterStatus messages.

  6. Avoid harmless Valgrind no-buffer-pin errors.

  7. Restructure libpq code to remove some duplicity

  8. libpq: Add target_session_attrs parameter.