Date: Fri, 29 May 2020 15:04:54 +0200
From: Tomas Vondra <tomas.vondra@2ndquadrant.com>
To: Jeff Davis <pgsql@j-davis.com>
Cc: Melanie Plageman <melanieplageman@gmail.com>,
	Thomas Munro <thomas.munro@gmail.com>,
	pgsql-hackers <pgsql-hackers@postgresql.org>,
	HeikkiLinnakangas <hlinnaka@iki.fi>
Subject: Re: Trouble with hashagg spill I/O pattern and costing
Message-ID: <20200529130454.el6qyuukj3ofs4eu@development>
References: <20200525225909.aifxjm53hw4xyzxd@development>
 <CA+hUKG+q3Ma+sTKn2316Ofof6u21hh0cGJHcJ-fAnLxw=LXevw@mail.gmail.com>
 <20200526141524.ohtvrp23vxosgivl@development>
 <e913544f6c236d37e8b0597cabd086570b5806f2.camel@j-davis.com>
 <20200526191511.d7pwz4awrvdm4p6j@development>
 <0589f869969de2f3e9713689159c7c4fbe1fcfa5.camel@j-davis.com>
 <CAAKRu_YNGQ-zw9rpXGtYhURMqdPk+HnMa+Wj-tnNx9cwENgDxg@mail.gmail.com>
 <20200527090704.onkjogxeg34ib5cn@development>
 <20200528185750.aamptuqdzrggqi5c@development>
 <0fb3456095d469cc20a9ee7e76edfe9237dae24f.camel@j-davis.com>
MIME-Version: 1.0
Content-Type: multipart/mixed; boundary="aldvszp5mqw37ch7"
Content-Disposition: inline
In-Reply-To: <0fb3456095d469cc20a9ee7e76edfe9237dae24f.camel@j-davis.com>
Precedence: bulk


--aldvszp5mqw37ch7
Content-Type: text/plain; charset=us-ascii; format=flowed
Content-Disposition: inline

On Thu, May 28, 2020 at 06:14:55PM -0700, Jeff Davis wrote:
>On Thu, 2020-05-28 at 20:57 +0200, Tomas Vondra wrote:
>> Attached is a patch adding CP_SMALL_TLIST to create_agg_plan and
>> create_groupingsets_plan.
>
>Looks good, except one question:
>
>Why would aggstrategy ever be MIXED when in create_agg_path? Wouldn't
>that only happen in create_groupingsets_path?
>

Ah, right. Yeah, we only need to check for AGG_HASH here. Moreover,
AGG_MIXED probably does not need the tlist tweak, because the input
should be pre-sorted as with AGG_SORTED.

And we should probably do similar check in the create_groupinsets_path,
I guess. At first I thought we can't do that before inspecting rollups,
which only happens later in the function, but now I see there's
aggstrategy in GroupingSetsPath too.


regards

-- 
Tomas Vondra                  http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

--aldvszp5mqw37ch7
Content-Type: text/plain; charset=us-ascii
Content-Disposition: attachment; filename="hashagg-tlist-fix-v2.patch"

diff --git a/src/backend/optimizer/plan/createplan.c b/src/backend/optimizer/plan/createplan.c
index 9941dfe65e..eab57fd74b 100644
--- a/src/backend/optimizer/plan/createplan.c
+++ b/src/backend/optimizer/plan/createplan.c
@@ -2113,12 +2113,21 @@ create_agg_plan(PlannerInfo *root, AggPath *best_path)
 	Plan	   *subplan;
 	List	   *tlist;
 	List	   *quals;
+	int			flags;
 
 	/*
 	 * Agg can project, so no need to be terribly picky about child tlist, but
-	 * we do need grouping columns to be available
+	 * we do need grouping columns to be available. We are a bit more careful
+	 * with hash aggregate, where we explicitly request small tlist to minimize
+	 * I/O needed for spilling (possibly not known already).
 	 */
-	subplan = create_plan_recurse(root, best_path->subpath, CP_LABEL_TLIST);
+	flags = CP_LABEL_TLIST;
+
+	/* ensure small tlist for hash aggregate */
+	if (best_path->aggstrategy == AGG_HASHED)
+		flags |= CP_SMALL_TLIST;
+
+	subplan = create_plan_recurse(root, best_path->subpath, flags);
 
 	tlist = build_path_tlist(root, &best_path->path);
 
@@ -2200,6 +2209,7 @@ create_groupingsets_plan(PlannerInfo *root, GroupingSetsPath *best_path)
 	int			maxref;
 	List	   *chain;
 	ListCell   *lc;
+	int			flags;
 
 	/* Shouldn't get here without grouping sets */
 	Assert(root->parse->groupingSets);
@@ -2207,9 +2217,17 @@ create_groupingsets_plan(PlannerInfo *root, GroupingSetsPath *best_path)
 
 	/*
 	 * Agg can project, so no need to be terribly picky about child tlist, but
-	 * we do need grouping columns to be available
+	 * we do need grouping columns to be available. We are a bit more careful
+	 * with hash aggregate, where we explicitly request small tlist to minimize
+	 * I/O needed for spilling (possibly not known already).
 	 */
-	subplan = create_plan_recurse(root, best_path->subpath, CP_LABEL_TLIST);
+	flags = CP_LABEL_TLIST;
+
+	/* ensure small tlist for hash aggregate */
+	if (best_path->aggstrategy == AGG_HASHED)
+		flags |= CP_SMALL_TLIST;
+
+	subplan = create_plan_recurse(root, best_path->subpath, flags);
 
 	/*
 	 * Compute the mapping from tleSortGroupRef to column index in the child's

--aldvszp5mqw37ch7--