From 5ecbf41a188f468331301da3685696b58120b85b Mon Sep 17 00:00:00 2001 From: Melanie Plageman Date: Wed, 11 Nov 2026 17:11:18 -0500 Subject: [PATCH v2 1/2] Avoid unreliable VACUUM reltuples estimates VACUUM protects pg_class.reltuples from tiny, non-random samples by retaining the old estimate when it scans very few pages. It did so even when the relation size changed, preserving a stale absolute count instead of scaling the old tuple density to the new size. VACUUM also reported an unknown live-tuple estimate as zero to the cumulative statistics system, overwriting its existing n_live_tup estimate. Commit b46e1e54d ("Allow on-access pruning to set pages all-visible") made these pre-existing problems reachable in ordinary workloads. A read can now mark most pages all-visible after INSERT, COPY, or CREATE TABLE AS but before VACUUM or ANALYZE establishes a tuple-density estimate. VACUUM may consequently skip nearly all populated pages and scan only a tiny, non-random tail sample while reltuples is still unknown or reflects a previously empty table. When no prior density is available, treat a sample of at most one page or less than two percent of the relation as insufficient, and leave reltuples unknown rather than extrapolating. Apply the small-sample stability rule only when the relation size is unchanged; otherwise, scale the old density to the new size. Pass an unknown estimate to the cumulative statistics system as -1, and preserve its existing n_live_tup estimate rather than replacing it with zero. --- src/backend/commands/vacuum.c | 30 +++++++++++++------- src/backend/utils/activity/pgstat_relation.c | 4 ++- 2 files changed, 22 insertions(+), 12 deletions(-) diff --git a/src/backend/commands/vacuum.c b/src/backend/commands/vacuum.c index 3a5c1efd6ac..c46de3c1a86 100644 --- a/src/backend/commands/vacuum.c +++ b/src/backend/commands/vacuum.c @@ -1371,6 +1371,20 @@ vac_estimate_reltuples(Relation relation, if (scanned_pages >= total_pages) return scanned_tuples; + /* + * With no prior density, a tiny non-random sample is not sufficient to + * estimate the relation's tuple count. This can happen when on-access + * pruning marks most pages all-visible before the first VACUUM. + */ + if (old_rel_tuples < 0 || old_rel_pages == 0) + { + if (scanned_pages <= 1 || + scanned_pages < (double) total_pages * 0.02) + return -1; + + return floor((scanned_tuples / scanned_pages) * total_pages + 0.5); + } + /* * When successive VACUUM commands scan the same few pages again and * again, without anything from the table really changing, there is a risk @@ -1381,23 +1395,17 @@ vac_estimate_reltuples(Relation relation, * If the relation is _exactly_ the same size according to the existing * pg_class entry, and only a few of its pages (less than 2%) were * scanned, keep the existing value of reltuples. Also keep the existing - * value when only a subset of rel's pages <= a single page were scanned. + * value when at most one page was scanned. When the relation size has + * changed, use the old density to account for the added or removed pages + * instead of preserving a stale tuple count. * * (Note: we might be returning -1 here.) */ if (old_rel_pages == total_pages && - scanned_pages < (double) total_pages * 0.02) - return old_rel_tuples; - if (scanned_pages <= 1) + (scanned_pages <= 1 || + scanned_pages < (double) total_pages * 0.02)) return old_rel_tuples; - /* - * If old density is unknown, we can't do much except scale up - * scanned_tuples to match total_pages. - */ - if (old_rel_tuples < 0 || old_rel_pages == 0) - return floor((scanned_tuples / scanned_pages) * total_pages + 0.5); - /* * Okay, we've covered the corner cases. The normal calculation is to * convert the old measurement to a density (tuples per page), then diff --git a/src/backend/utils/activity/pgstat_relation.c b/src/backend/utils/activity/pgstat_relation.c index 5c70543ba88..4be20826cd1 100644 --- a/src/backend/utils/activity/pgstat_relation.c +++ b/src/backend/utils/activity/pgstat_relation.c @@ -258,6 +258,7 @@ pgstat_drop_relation(Relation rel) /* * Report that the table was just vacuumed and flush IO statistics. + * livetuples is -1 if VACUUM could not produce a new estimate. */ void pgstat_report_vacuum(Relation rel, PgStat_Counter livetuples, @@ -284,7 +285,8 @@ pgstat_report_vacuum(Relation rel, PgStat_Counter livetuples, shtabentry = (PgStatShared_Relation *) entry_ref->shared_stats; tabentry = &shtabentry->stats; - tabentry->live_tuples = livetuples; + if (livetuples >= 0) + tabentry->live_tuples = livetuples; tabentry->dead_tuples = deadtuples; /* -- 2.43.0