Обсуждение: BUG #19030: Hash join leads to extremely high memory usage

The following bug has been logged on the website:

Bug reference:      19030
Logged by:          Marc-Olaf Jaschke
Email address:      moj@dshare.de
PostgreSQL version: 17.6
Operating system:   Linux
Description:

Description
- Two tables, left join
- The left table has significantly fewer rows than the right table
- The left table has very large rows (many columns with high memory usage)
- The left table has many null values in the join column
- A hash join is used
- The hash node is built from the left table
- The query results in extremely high memory usage (100x work_mem in the
example, > 1000x in real case)
- Reliably causing a PostgreSQL server to crash in production
- With enable_hashjoin = false, the query runs without any issues

Example
- Simplified artificial example – but I hope it simulates a real problem on
a production system well.
- Simulate large rows with one big column
- Running the newest version with mostly default settings


==========================================


select version();
-- > PostgreSQL 17.6 (Debian 17.6-1.pgdg13+1) on aarch64-unknown-linux-gnu,
compiled by gcc (Debian 14.2.0-19) 14.2.0, 64-bit

show work_mem;
-- > 4MB

show hash_mem_multiplier;
-- > 2

-- speed up the test case
set default_toast_compression = lz4;

-- show high memory usage without crashing the server
set max_parallel_workers_per_gather = 0;


create table left_ as
select
case when i% 2 = 0 then i::text end c,
repeat('x', 10_000) big
from
generate_series(1, 10_000_000) i;

create table right_ as
select
case when i% 2 = 0 then i::text end c
from
generate_series(1, 20_000_000) i;

analyze left_, right_;


explain (analyze, memory)
select
*
from
left_
natural left join
right_;

-- > Buckets: 131072 (originally 131072)  Batches: 131072 (originally 256)
Memory Usage: 440482kB

PG Bug reporting form <noreply@postgresql.org> writes:
> Description
> - Two tables, left join
> - The left table has significantly fewer rows than the right table
> - The left table has very large rows (many columns with high memory usage)
> - The left table has many null values in the join column
> - A hash join is used
> - The hash node is built from the left table
> - The query results in extremely high memory usage (100x work_mem in the
> example, > 1000x in real case)

If this is specific to the case of many null join values, it's a known
problem that I have a patch in the queue for [1].

On your example, I get this on HEAD:


                                                             QUERY PLAN
            

------------------------------------------------------------------------------------------------------------------------------------
 Hash Right Join  (cost=459884.00..1367112.00 rows=10000000 width=65) (actual time=2557.730..11727.429 rows=10000000.00
loops=1)
   Hash Cond: (right_.c = left_.c)
   Buffers: shared read=201216, temp read=131098 written=131098
   ->  Seq Scan on right_  (cost=0.00..283520.00 rows=20000000 width=8) (actual time=0.060..806.467 rows=20000000.00
loops=1)
         Buffers: shared read=83520
   ->  Hash  (cost=217696.00..217696.00 rows=10000000 width=65) (actual time=2494.680..2494.680 rows=10000000.00
loops=1)
         Buckets: 131072 (originally 131072)  Batches: 16384 (originally 256)  Memory Usage: 440509kB
         Buffers: shared read=117696, temp written=46737
         ->  Seq Scan on left_  (cost=0.00..217696.00 rows=10000000 width=65) (actual time=0.096..481.139
rows=10000000.00loops=1) 
               Buffers: shared read=117696
 Planning:
   Buffers: shared hit=136 read=35
   Memory: used=21kB  allocated=32kB
 Planning Time: 0.703 ms
 Execution Time: 11946.100 ms
(15 rows)

and this with the aforesaid patch:

                                                             QUERY PLAN
            

------------------------------------------------------------------------------------------------------------------------------------
 Hash Right Join  (cost=459884.00..1367112.00 rows=10000000 width=65) (actual time=1553.342..6589.352 rows=10000000.00
loops=1)
   Hash Cond: (right_.c = left_.c)
   Buffers: shared hit=188 read=201028, temp read=128626 written=128626
   ->  Seq Scan on right_  (cost=0.00..283520.00 rows=20000000 width=8) (actual time=0.099..750.684 rows=20000000.00
loops=1)
         Buffers: shared hit=94 read=83426
   ->  Hash  (cost=217696.00..217696.00 rows=10000000 width=65) (actual time=1551.739..1551.740 rows=5000000.00
loops=1)
         Buckets: 131072  Batches: 256  Memory Usage: 2906kB
         Buffers: shared hit=94 read=117602, temp written=93662
         ->  Seq Scan on left_  (cost=0.00..217696.00 rows=10000000 width=65) (actual time=0.106..470.671
rows=10000000.00loops=1) 
               Buffers: shared hit=94 read=117602
 Planning:
   Memory: used=20kB  allocated=32kB
 Planning Time: 0.122 ms
 Execution Time: 6827.197 ms
(14 rows)

(Hmm, looking at this, it's apparent that the patch causes the
Hash node not to count the null-keyed rows in its EXPLAIN output.
On the one hand, that's an accurate reflection of how much went into
the hash table, but on the other hand it's pretty confusing.)

I doubt we'd consider back-patching such a change, but if you want
to see it happen for v19, you could help by reviewing/testing.

            regards, tom lane

[1] https://www.postgresql.org/message-id/flat/3061845.1746486714%40sss.pgh.pa.us