Jan Wieck <jan@wi3ck.info> wrote:
>>> The patch uses a heuristic method of detecting when the hash table
>>> should be destroyed and recreated. InvalidateConstraintCacheCallBack()
>>> adds the current size of the hash table to a counter. When that sum
>>> reaches 1,000,000, the hash table is flushed. This improves the schema
>>> restore of a database with 100,000 foreign keys by factor 3.
>>>
>>> According to my tests the patch does not interfere with the bulk
>>> updates, the original feature was supposed to improve.
>>
>> In the real-world customer case that caused you to look into this,
>> I thought 45ba424f drove schema-only restore time from 2 hours to
>> about 25 hours, and that this patch takes it back down to 2 hours.
>> Am I remembering right? And this came about because it added
>> 20-some hours to a pg_upgrade run?
>
> From 2 hours to >50, but yes, this is that case.
>
>> If there are no objections, I will push this as a bug fix back to
>> 9.3, where the performance regression was introduced.
Hearing no objections, pushed with one minor tweak Jan and I
discussed off-list -- the original patch duplicated a bit of code
that we agreed should be split into a static function and called
from both places, to protect against someone later changing one and
missing the other.
--
Kevin Grittner
EDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
