Date: Fri, 23 Oct 2020 17:44:51 +0200
From: Jehan-Guillaume de Rorthais <jgdr@dalibo.com>
To: PostgreSQL-development <pgsql-hackers@postgresql.org>, Peter Geoghegan
 <pg@bowt.ie>
Subject: vacuum -vs reltuples on insert only index
Message-ID: <20201023174451.69e358f1@firost>
Organization: Dalibo
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Precedence: bulk

Hello,

I've found a behavior change with pg_class.reltuples on btree index. With only
insert activity on a table, when an index is processed, its related reltuples
is set to 0. Here is a demo script:

  -- force index cleanup
  set vacuum_cleanup_index_scale_factor to 0;

  drop table if exists t;
  create table t as select i from generate_series(1, 100) i;
  create index t_i on t(i);

  -- after index creation its reltuples is correct
  select reltuples from pg_class where relname = 't_i' 
  -- result: reltuples | 100

  -- vacuum set index reltuples to 0
  vacuum t;
  select reltuples from pg_class where relname = 't_i' 
  -- result: reltuples | 0

  -- analyze set it back to correct value
  analyze t;
  select reltuples from pg_class where relname = 't_i' 
  -- result: reltuples | 100

  -- insert + vacuum reset it again to 0
  insert into t values(101);
  vacuum (verbose off, analyze on, index_cleanup on) t;
  select reltuples from pg_class where relname = 't_i' 
  -- result: reltuples | 0

  -- delete + vacuum set it back to correct value
  delete from t where i=10;
  vacuum (verbose off, analyze on, index_cleanup on) t;
  select reltuples from pg_class where relname = 't_i' 
  -- result: reltuples | 100

  -- and back to 0 again with insert+vacuum
  insert into t values(102);
  vacuum (verbose off, analyze on, index_cleanup on) t;
  select reltuples from pg_class where relname = 't_i' 
  -- result: reltuples | 0

Before 0d861bbb70, btvacuumpage was adding to relation stats the number of
leaving lines in the block using:

  stats->num_index_tuples += maxoff - minoff + 1;

After 0d861bbb70, it is set using new variable nhtidslive:

  stats->num_index_tuples += nhtidslive

However, nhtidslive is only incremented if callback (IndexBulkDeleteCallback)
is set, which seems not to be the case on select-only workload.

A naive fix might be to use "maxoff - minoff + 1" when callback==NULL.

Thoughts?

Regards,