Query plan getting less efficient over time with frequent updates and deletes..

Hi all: we have a table which receives frequent daily updates and deletes
on the order of 100-600k. The overall row length is approximately 80m. This
table has 50 indexes and 303 columns and is quite frequently queried by
humans and applications.

I've been in the habit of using pg_repack maybe once a month on this table
because I can't quite figure out why querying gets bogged down. The vacuum
and analyze thresholds are set such that the table is both auto vacuumed
and analyzed every night.

My questions/concerns:

- this gets vacuumed and analyzed due to auto thresholds every night, so
it's surprising that the planner's choices would change and choose a worse
plan right?
- I look at pg_stat_user_tables to asses n_dead_tup along with n_tup_del
and n_tup_upd to try and assess if there is table bloat and this seems to
be under control
- I look at pgstattuple to check free_space and try and guess how much is
being given back to the server for new rows, it's roughly 20-28% and it
seems OK?
- what other factors are at play here? index bloat? disk IO? etc? something
I can eyeball better in pg_stat_user_tables, or pgstattuple?

How can I make this table behave properly, with all of its updates, and
without restoring to frequent pg_repack?

The much faster version of the query plan does a bitmap heap scan with a
bitmap index scan. The slower version just does an index scan on the same
index.

Here are the EXPLAIN ANALYZE VERBOSE results of a routine SELECT query with
a few columns in a WHERE clause. I've abbreviated these otherwise it would
just be too much text, glad to share more. After a fresh pg_repack, the
former is performed for a while (like two weeks) then it reverts to the
latter and everyone is unhappy.

Fast, better!

Sort (cost=20279.85..20281.77 rows=768 width=380) (actual
time=218.563..219.431 rows=7892 loops=1)
Output: p.season, p.game_date, ... p.pitch_seq
Sort Key: p.game_date DESC, p.game_bam_id, ((p.at_bat_index + 1)),
p.pitch_seq
Sort Method: quicksort Memory: 4215kB
-> Bitmap Heap Scan on stats.pitches p (cost=58.49..20243.04 rows=768
width=380) (actual time=11.319..167.570 rows=7892 loops=1)
Output: p.season, p.game_date, ... p.pitch_seq
Recheck Cond: (p.batter_common_key = 2181560)
Filter: ((p.pitch_type IS NOT NULL) AND p.is_pitch AND
(COALESCE(p.pitch_code, p.pitch_result) IS NOT NULL) AND (p.season = ANY
('{2024,2023,2022}'::integer[])) AND (p.pitch_type <> ALL
('{UN,XX,PO,IB,AB}'::text[])) AND (p.game_type = ANY
('{R,F,D,L,W,C,S}'::text[])))
Rows Removed by Filter: 2824
Heap Blocks: exact=10327
-> Bitmap Index Scan on pitches_batter_common_key_idx
(cost=0.00..58.26 rows=5026 width=0) (actual time=6.625..6.625 rows=10716
loops=1)
Index Cond: (p.batter_common_key = 2181560)
Query Identifier: -9080545043453138058
Planning Time: 3.358 ms
Execution Time: 224.082 ms

Slow, bad!

Sort (cost=20381.86..20383.74 rows=753 width=381) (actual
time=6344.784..6345.568 rows=7892 loops=1)
Output: p.season, p.game_date, ... p.pitch_seq
Sort Key: p.game_date DESC, p.game_bam_id, ((p.at_bat_index + 1)),
p.pitch_seq
Sort Method: quicksort Memory: 4215kB
-> Index Scan using pitches_batter_common_key_idx on stats.pitches p
(cost=0.60..20345.88 rows=753 width=381) (actual time=6.765..6294.745
rows=7892 loops=1)
Output: p.season, p.game_date, ... p.pitch_seq
Index Cond: (p.batter_common_key = 2181560)
Filter: ((p.pitch_type IS NOT NULL) AND p.is_pitch AND
(COALESCE(p.pitch_code, p.pitch_result) IS NOT NULL) AND (p.season = ANY
('{2024,2023,2022}'::integer[])) AND (p.pitch_type <> ALL
('{UN,XX,PO,IB,AB}'::text[])) AND (p.game_type = ANY
('{R,F,D,L,W,C,S}'::text[])))
Rows Removed by Filter: 2824
Query Identifier: -9080545043453138058
Planning Time: 3.197 ms
Execution Time: 6349.608 ms

--
Wells Oliver
wells(dot)oliver(at)gmail(dot)com <wellsoliver(at)gmail(dot)com>