On 3/5/2025 18:52, Sergey Koposov wrote:
> On Sat, 2025-05-03 at 12:27 -0400, Tom Lane wrote:
>> Yeah, this confirms the idea that the hashtable has exploded into an
>> unreasonable number of buckets and batches. I don't know why a
>> parallel hash join would be more prone to do that than a non-parallel
>> one, though. I'm hoping some of the folks who worked on PHJ will
>> look at this.
> Here're my memory settings
>
> shared_buffers = 32GB
> work_mem = 1GB
May you show statistics from the pg_statistic table on two columns:
"d.objid" and "gaps1.original_ext_source_id"? At least stanullfrac,
stadistinct, stawidth and MCV frequencies.
Also, an EXPLAIN ANALYZE on this query, which successfully finished
execution without parallel workers, might provide quick insights.
--
regards, Andrei Lepikhov
