citus/src
Jelte Fennema 4c3934272f
Improve performance of citus_shards (#5036)
We were effectively joining on a calculated column because of our calls
to `shard_name`. This caused a really bad plan to be generated. In my
specific case it was taking ~18 seconds to show the output of
citus_shards. It had this explain plan:

```
                                                                                                       QUERY PLAN
─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
 Subquery Scan on citus_shards  (cost=18369.74..18437.34 rows=5408 width=124) (actual time=18277.461..18278.509 rows=5408 loops=1)
   ->  Sort  (cost=18369.74..18383.26 rows=5408 width=156) (actual time=18277.457..18277.726 rows=5408 loops=1)
         Sort Key: ((pg_dist_shard.logicalrelid)::text), pg_dist_shard.shardid
         Sort Method: quicksort  Memory: 1629kB
         CTE shard_sizes
           ->  Function Scan on citus_shard_sizes  (cost=0.00..10.00 rows=1000 width=40) (actual time=71.137..71.934 rows=5413 loops=1)
         ->  Hash Join  (cost=177.62..18024.42 rows=5408 width=156) (actual time=77.985..18257.237 rows=5408 loops=1)
               Hash Cond: ((pg_dist_shard.logicalrelid)::oid = (pg_dist_partition.logicalrelid)::oid)
               ->  Hash Join  (cost=169.81..371.98 rows=5408 width=48) (actual time=1.415..13.166 rows=5408 loops=1)
                     Hash Cond: (pg_dist_placement.groupid = pg_dist_node.groupid)
                     ->  Hash Join  (cost=168.68..296.49 rows=5408 width=16) (actual time=1.403..10.011 rows=5408 loops=1)
                           Hash Cond: (pg_dist_placement.shardid = pg_dist_shard.shardid)
                           ->  Seq Scan on pg_dist_placement  (cost=0.00..113.60 rows=5408 width=12) (actual time=0.004..3.684 rows=5408 loops=1)
                                 Filter: (shardstate = 1)
                           ->  Hash  (cost=101.08..101.08 rows=5408 width=12) (actual time=1.385..1.386 rows=5408 loops=1)
                                 Buckets: 8192  Batches: 1  Memory Usage: 318kB
                                 ->  Seq Scan on pg_dist_shard  (cost=0.00..101.08 rows=5408 width=12) (actual time=0.003..0.688 rows=5408 loops=1)
                     ->  Hash  (cost=1.06..1.06 rows=6 width=40) (actual time=0.007..0.007 rows=6 loops=1)
                           Buckets: 1024  Batches: 1  Memory Usage: 9kB
                           ->  Seq Scan on pg_dist_node  (cost=0.00..1.06 rows=6 width=40) (actual time=0.004..0.005 rows=6 loops=1)
               ->  Hash  (cost=5.69..5.69 rows=169 width=130) (actual time=0.070..0.071 rows=169 loops=1)
                     Buckets: 1024  Batches: 1  Memory Usage: 36kB
                     ->  Seq Scan on pg_dist_partition  (cost=0.00..5.69 rows=169 width=130) (actual time=0.009..0.041 rows=169 loops=1)
               SubPlan 2
                 ->  Limit  (cost=0.00..3.25 rows=1 width=8) (actual time=3.370..3.370 rows=1 loops=5408)
                       ->  CTE Scan on shard_sizes  (cost=0.00..32.50 rows=10 width=8) (actual time=3.369..3.369 rows=1 loops=5408)
                             Filter: ((shard_name(pg_dist_shard.logicalrelid, pg_dist_shard.shardid) = table_name) OR (('public.'::text || shard_name(pg_dist_shard.logicalrelid, pg_dist_shard.shardid)) = table_name))
                             Rows Removed by Filter: 2707
 Planning Time: 0.705 ms
 Execution Time: 18278.877 ms
```

With the changes it only takes 180ms to show the same output:
```
                                                                              QUERY PLAN
──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
 Sort  (cost=904.59..918.11 rows=5408 width=156) (actual time=182.508..182.960 rows=5408 loops=1)
   Sort Key: ((pg_dist_shard.logicalrelid)::text), pg_dist_shard.shardid
   Sort Method: quicksort  Memory: 1629kB
   ->  Hash Join  (cost=418.03..569.27 rows=5408 width=156) (actual time=136.333..146.591 rows=5408 loops=1)
         Hash Cond: ((pg_dist_shard.logicalrelid)::oid = (pg_dist_partition.logicalrelid)::oid)
         ->  Hash Join  (cost=410.22..492.83 rows=5408 width=56) (actual time=136.231..140.132 rows=5408 loops=1)
               Hash Cond: (pg_dist_placement.groupid = pg_dist_node.groupid)
               ->  Hash Right Join  (cost=409.09..417.34 rows=5408 width=24) (actual time=136.218..138.890 rows=5408 loops=1)
                     Hash Cond: ((((regexp_matches(citus_shard_sizes.table_name, '_(\d+)$'::text))[1])::integer) = pg_dist_shard.shardid)
                     ->  HashAggregate  (cost=45.00..48.50 rows=200 width=12) (actual time=131.609..132.481 rows=5408 loops=1)
                           Group Key: ((regexp_matches(citus_shard_sizes.table_name, '_(\d+)$'::text))[1])::integer
                           Batches: 1  Memory Usage: 737kB
                           ->  Result  (cost=0.00..40.00 rows=1000 width=12) (actual time=107.786..129.831 rows=5408 loops=1)
                                 ->  ProjectSet  (cost=0.00..22.50 rows=1000 width=40) (actual time=107.780..128.492 rows=5408 loops=1)
                                       ->  Function Scan on citus_shard_sizes  (cost=0.00..10.00 rows=1000 width=40) (actual time=107.746..108.107 rows=5414 loops=1)
                     ->  Hash  (cost=296.49..296.49 rows=5408 width=16) (actual time=4.595..4.598 rows=5408 loops=1)
                           Buckets: 8192  Batches: 1  Memory Usage: 339kB
                           ->  Hash Join  (cost=168.68..296.49 rows=5408 width=16) (actual time=1.702..3.783 rows=5408 loops=1)
                                 Hash Cond: (pg_dist_placement.shardid = pg_dist_shard.shardid)
                                 ->  Seq Scan on pg_dist_placement  (cost=0.00..113.60 rows=5408 width=12) (actual time=0.004..0.837 rows=5408 loops=1)
                                       Filter: (shardstate = 1)
                                 ->  Hash  (cost=101.08..101.08 rows=5408 width=12) (actual time=1.683..1.685 rows=5408 loops=1)
                                       Buckets: 8192  Batches: 1  Memory Usage: 318kB
                                       ->  Seq Scan on pg_dist_shard  (cost=0.00..101.08 rows=5408 width=12) (actual time=0.004..0.824 rows=5408 loops=1)
               ->  Hash  (cost=1.06..1.06 rows=6 width=40) (actual time=0.007..0.008 rows=6 loops=1)
                     Buckets: 1024  Batches: 1  Memory Usage: 9kB
                     ->  Seq Scan on pg_dist_node  (cost=0.00..1.06 rows=6 width=40) (actual time=0.004..0.006 rows=6 loops=1)
         ->  Hash  (cost=5.69..5.69 rows=169 width=130) (actual time=0.079..0.079 rows=169 loops=1)
               Buckets: 1024  Batches: 1  Memory Usage: 36kB
               ->  Seq Scan on pg_dist_partition  (cost=0.00..5.69 rows=169 width=130) (actual time=0.011..0.046 rows=169 loops=1)
 Planning Time: 0.789 ms
 Execution Time: 184.095 ms
 ```
2021-06-14 13:32:30 +02:00
..
backend Improve performance of citus_shards (#5036) 2021-06-14 13:32:30 +02:00
include Do not rely on fk cache when truncating local data (#5018) 2021-06-07 11:56:48 +03:00
test Enforce table opt constraints when using alter_columnar_table_set (#5029) 2021-06-08 17:39:16 +03:00