Description

Elasticsearch's nightly benchmarks noticed a strange regression where the entire merge time for one of our vector benchmarks almost doubled in size. (Note the cumulative merge time graphed here: https://elasticsearch-benchmarks.elastic.co/#tracks/so_vector/nightly/default/90d)

After some digging and testing (manually running different lucene patches on the same hardware), I verified it was due to #13907

I am fairly ignorant of the impact of this change, nor what the expected impact should be.

Maybe Elasticsearch is providing poor settings to Lucene and didn't handle the change correctly?

Or the change itself made a larger impact that originally designed?

Version and environment details

Lucene 10.1

Intel IceLake CPU

8 vCPUs, 16GB RAM and 1x300GiB SSD disk