MIME-Version: 1.0
References: 
 <CAMbWs48jzLrPt1J_00ZcPZXWUQKawQOFE8ROc-ADiYqsqrpBNw@mail.gmail.com>
 <87il22cj51.fsf@163.com>
 <CAMbWs49=eAd2W9jCtGhaZPPp+SOC_2rg16RTG74xAht=hkr5JQ@mail.gmail.com>
 <CAMbWs49Nc4M3H+eCf1+8w8piDyEECjRb-gK_JMF4VvcyWwGEVQ@mail.gmail.com>
 <CAMbWs49E_dR0nobsExsyetpnBpHObLTsQLsEbWKQLkh0omPxNg@mail.gmail.com>
 <CAMbWs49B_qUiHvu2EqLHZRpLr3p_+QPBs50n2=L5ibYzniwTzA@mail.gmail.com>
 <CAMbWs48KCQtDymnYi4M=Vz+WMzo3fkBxffJsyk6VX6hOXXv+VA@mail.gmail.com>
 <CAMbWs49sv_MuOYqqrtmBN_oYf8VSQ2BXDwXaTpJTn_YfwyYdWQ@mail.gmail.com>
 <CAMbWs49U8Sddx_fGszPdvA3jp_nheynxaqm5Y4NqMV21VBYAuQ@mail.gmail.com>
 <CAMbWs4-LwyOg9ga+NVF7yQbMi0ZsZdN1G_sO2v=YJHV18=19+A@mail.gmail.com>
 <CALA8mJquG_zCJXfVwash5LKqHGtZXQmq7RfTSaRDUzGYeW=7Rw@mail.gmail.com>
 <CAMbWs4_EjgcBib5+y1LYcGB3EK3Y6R+OOxGKfJo42fDovadk1g@mail.gmail.com>
 <CALA8mJqe0anNM8_V6cOeOQnCHUTQggn7iOQNyQr1VaN_xMjz+w@mail.gmail.com>
 <CAMbWs48eE-s-jCicC8pSVfXk8Ws-ZvUKnsw8qH-DkVBdYv0eJQ@mail.gmail.com>
 <CAMbWs483a7-8M0pDttG44r-+8Gevn9VG0xNceE3WpkEQxJXPZw@mail.gmail.com>
 <CAHewXNmYM6DvR_kaxDL0w0fz9BwKbac+TSU3QS10aA3cXHyMmA@mail.gmail.com>
 <CA+TgmoaxH=P63hLYgyJJcEbMRnw3xi16d=HxFi1j-m7MhH6W_w@mail.gmail.com>
 <CAMbWs4_cOnpGsywj9Jt1WAgzJLW9Rxt5X13cfGz4iN2qvZQ68g@mail.gmail.com>
 <CA+Tgmob0q7bRbsFTVDMjxHE6zA4uDQLQa-s0CtwUw49V53UL_A@mail.gmail.com>
 <CAMbWs4-Xru_eKBeRHFduigSGihdixFWVTR8A+dtMw7Mao+RkJA@mail.gmail.com>
 <CAMbWs49dLjSSQRWeud+KSN0G531ciZdYoLBd5qktXA+3JQm_UQ@mail.gmail.com>
 <CAMbWs48LXGC-Y63YtzEeM-3f0NUXWCUEMs7XwGzywXTjUNMcxQ@mail.gmail.com>
 <CAMbWs48XdzvnwfTHWxQ7qK-yjvdrbwsPpqhJBuKDnO+hcbsVwA@mail.gmail.com>
 <CA+TgmoaO-7RHdyJuizWChXZm7EJGvDcfoePDDEyUA-y8vTB1tg@mail.gmail.com>
 <CAMbWs4-+jXRpKuFMZa08bS34-TBka3qqjVMAUjF=-1RA9BKvgg@mail.gmail.com>
 <CA+TgmoZapU1y59-s3o8oPt7Hv+cxRh_34FMu6MXumomLe+U1Cw@mail.gmail.com>
 <CAMbWs4_sEeeBmucBzbamBMfA9uLxVmOc_MV=ZpSyDbTcrUO_XQ@mail.gmail.com>
 <CA+Tgmob4fnv57PQB0Oox86mHSJQ0vVL249eT=gqPvrMkG7h1zw@mail.gmail.com>
 <CAMbWs489NYyTcCTbrUi7hPXKtNY5vHrrFcHyMRAv=CA5WsszVw@mail.gmail.com>
 <CA+TgmoazmDdcc7NeTo3WM5HW3DASNP4rfZw6X+2nnQKHampOng@mail.gmail.com>
 <CAMbWs49bYr-ULhA+-At0iQ+NaFKy72AWB6jzughk8MPTiY+gMQ@mail.gmail.com>
 <CA+TgmoYa-zexdbc5nO_D6oxPMZYs06hkYwZK5Dufq+4Hhe6uNQ@mail.gmail.com>
 <CAMbWs4_aji0kME490phz6nTXnPToddUn19OF3rLm1g4TbNkuzQ@mail.gmail.com>
 <CA+Tgmoa3+G_=8XuQWN+0ugv6r-WV6ruFESpOxpXAAKrne3oVDQ@mail.gmail.com>
 <CAMbWs49qiox13EKb7bqgLu7Gu9oar+xe6KMwBjgFwod3JzPfUw@mail.gmail.com>
 <CAMbWs48F8WGA-Lzj1Dk76mFqRFxPEwG2_9Zb7+pFs8oi6ew2pw@mail.gmail.com>
 <CAMbWs484ms=WRZamOyWnVditREKFqipLsdaQjcv2uKur8SZuqw@mail.gmail.com>
 <CAMbWs49bL2ZMSc0W4G8=R7bjaa-vO6grucEOFYLZFUZE7+nzrQ@mail.gmail.com>
 <CA+TgmoaycfwRGtVpzmgUbEibP5zyuNRZoJ4Sc0475tRW4rofoQ@mail.gmail.com>
In-Reply-To: 
 <CA+TgmoaycfwRGtVpzmgUbEibP5zyuNRZoJ4Sc0475tRW4rofoQ@mail.gmail.com>
From: Richard Guo <guofenglinux@gmail.com>
Date: Tue, 9 Sep 2025 20:18:20 +0900
Message-ID: 
 <CAMbWs4-sEi3GeZA2trW87OMDuPhJQ6QhZ-KPMNFdCqC6BM+UMQ@mail.gmail.com>
Subject: Re: Eager aggregation, take 3
To: Robert Haas <robertmhaas@gmail.com>
Cc: Tom Lane <tgl@sss.pgh.pa.us>, Tender Wang <tndrwang@gmail.com>,
	Paul George <p.a.george19@gmail.com>, Andy Fan <zhihuifan1213@163.com>,
	PostgreSQL-development <pgsql-hackers@postgresql.org>,
 pgsql-hackers@lists.postgresql.org,
	Matheus Alcantara <matheusssilv97@gmail.com>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Archived-At: 
 <https://www.postgresql.org/message-id/CAMbWs4-sEi3GeZA2trW87OMDuPhJQ6QhZ-KPMNFdCqC6BM%2BUMQ%40mail.gmail.com>
Precedence: bulk

On Fri, Sep 5, 2025 at 11:50=E2=80=AFPM Robert Haas <robertmhaas@gmail.com>=
 wrote:
> Like Matheus, I think a GUC is reasonable. A significant danger here
> appears to be the possibility of a performance cliff, where queries
> are optimized very different when the ratio is 9.99 vs. 10.01, say. It
> would be nice if there were some way to mitigate that danger, but at
> least a GUC avoids chaining the performance of the whole system to a
> hard-coded value.

Yeah, I think the performance cliff issue does exist.  It might be
mitigated by carefully selecting the threshold value to ensure that
small differences in the average group size near the boundary don't
cause big performance swings with and without eager aggregation, but
this doesn't seem like an easy task.

How is this issue avoided in other thresholds?  For example, with
min_parallel_table_scan_size, is there a performance cliff when the
table size is 7.99MB vs. 8.01MB, where a parallel scan is considered
in the latter case but not the former?

> It might be worth considering whether there are heuristics other than
> the group size that could help here. Possibly that's just making
> things more complicated to no benefit. It seems to me, for example,
> that reducing 100 rows to 10 is quite different from reducing a
> million rows to 100,000. On the whole, the latter seems more likely to
> work out well, but it's tricky, because the effort expended per group
> can be arbitrarily high. I think we do want to let the cost model make
> most of the decisions, and just use this threshold to prune ideas that
> are obviously bad at an early stage. That said, it's worth thinking
> about how this interacts with the just-considered-one-eager-agg
> strategy. Does this threshold apply before or after that rule?

If I understand correctly, this means that we need to explore each
join level to find out the most optimal position for applying partial
aggregation.  For example, suppose Agg(B) reduces 100 rows to 10, and
Agg(A JOIN B) reduces a million rows to 100,000, it might be better to
apply partial aggregation at the (A JOIN B) level rather than just
over B.  However, that's not always the case: the Agg(B) option can
reduce the number of input rows to the join earlier, potentially
outperforming the Agg(A JOIN B) approach.  Therefore, we need to
consider both options and compare their costs.

This is actually what the patch used to do before I introduced the
always-push-to-lowest heuristic.

> For instance, consider AGG(FACT_TABLE JOIN DIMENSION_TABLE), like a
> count of orders grouped by customer name. Aggregating on the dimension
> table (in this case, the list of customers) is probably useless, but
> aggregating on the join column of the fact table has a good chance of
> being useful. If we consider only one of those strategies, we want it
> to be the right one. This threshold could be the thing that helps us
> to get it right.

Now I see what you meant.  However, in the current implementation, we
only push partial aggregation down to relations that contain all the
aggregation columns.  So, in the case you mentioned, if the
aggregation columns come from the dimension table, unfortunately, we
don't have the option to partially aggregate the fact table.

The paper does discuss several other transformations, such as "Eager
Count", "Double Eager", and "Eager Split", that can perform partial
aggregation on relations that don't contain aggregation columns, or
even on both sides of the join.  However, those are beyond the scope
of this patch.

- Richard