MIME-Version: 1.0
From: Sivaprasad <sivaprasad.off@gmail.com>
Date: Thu, 25 Jun 2026 07:16:07 +0530
Message-ID: 
 <CAHi-N2Fqs=7vUfNzDopsNnxB145kOz7FAOXofxw=gFSqaG_vrw@mail.gmail.com>
Subject: Pre-PG18 parallel scan estimation: accessing planned worker count
 from amestimateparallelscan()
To: pgsql-general@lists.postgresql.org
Content-Type: multipart/alternative; boundary="0000000000008ca95d06550a284a"
Archived-At: 
 <https://www.postgresql.org/message-id/CAHi-N2Fqs%3D7vUfNzDopsNnxB145kOz7FAOXofxw%3DgFSqaG_vrw%40mail.gmail.com>
Precedence: bulk

--0000000000008ca95d06550a284a
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

Hello,

While adding parallel scan support to the Biscuit
<https://github.com/CrystallineCore/Biscuit> index access method, I ran
into a question about the pre-PG18 parallel scan API.

Biscuit partitions scan work into a fixed number of segments at estimation
time and assigns those segments to parallel participants via a
shared-memory descriptor in the DSM segment. Because the segment layout
depends on the number of participants, we need the planned worker count
during amestimateparallelscan() =E2=80=94 the only callback that both sizes=
 the DSM
allocation and runs in the same process immediately before
aminitparallelscan().

The callback signatures across versions are:
-

   -

   PG16: Size (*)(void)
   -

   PG17: Size (*)(int nkeys, int norderbys)
   -

   PG18+: Size (*)(Relation, int *nworkers*, int nchunks)


In PG18+, nworkers is available directly. In PG16 and PG17 it is not.

My question: prior to PG18, was there any supported mechanism by which an
index AM could determine the planner's intended parallel worker count
during scan estimation?

Thanks,
Sivaprasad Murali
Biscuit index access method

--0000000000008ca95d06550a284a
Content-Type: text/html; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

<div dir=3D"ltr"><p>Hello,<br><br>While adding parallel scan support to the=
 <a href=3D"https://github.com/CrystallineCore/Biscuit">Biscuit</a> index a=
ccess method, I ran into a question about the pre-PG18 parallel scan API.<b=
r><br>Biscuit partitions scan work into a fixed number of segments at estim=
ation time and assigns those segments to parallel participants via a shared=
-memory descriptor in the DSM segment. Because the segment layout depends o=
n the number of participants, we need the planned worker count during amest=
imateparallelscan() =E2=80=94 the only callback that both sizes the DSM all=
ocation and runs in the same process immediately before aminitparallelscan(=
).<br><br>The callback signatures across versions are:</p><li><ul><li><p>PG=
16: <code>Size (*)(void)</code></p></li><li><p>PG17: <code>Size (*)(int nke=
ys, int norderbys)</code></p></li><li><p>PG18+: <code>Size (*)(Relation, in=
t <b>nworkers</b>, int nchunks)</code></p></li></ul></li><p><br>In PG18+, n=
workers is available directly. In PG16 and PG17 it is not.<br><br>My questi=
on: prior to PG18, was there any supported mechanism by which an index AM c=
ould determine the planner&#39;s intended parallel worker count during scan=
 estimation?=C2=A0<br><br>Thanks,<br>Sivaprasad Murali<br>Biscuit index acc=
ess method</p></div>

--0000000000008ca95d06550a284a--