This seems to have been caused by defaulting to the vectorized SQL execution engine for all queries in #55713, which was merged on December 3rd. The vectorize_row_count_threshold was subsequently removed in #62164.

Using commit 5766c37 from March 17th, I ran ycsb/E/nodes=3 benchmarks on GCE with DefaultVectorizeRowCountThreshold set to 0 and 1000, representative results below:

`DefaultVectorizeRowCountThreshold = 0`

_elapsed___errors_____ops(total)___ops/sec(cum)__avg(ms)__p50(ms)__p95(ms)__p99(ms)_pMax(ms)__total
  600.0s        0          24692           41.2     19.0     15.7     48.2     71.3    125.8  insert

_elapsed___errors_____ops(total)___ops/sec(cum)__avg(ms)__p50(ms)__p95(ms)__p99(ms)_pMax(ms)__total
  600.0s        0         467824          779.7    122.1     75.5    419.4    570.4   1006.6  scan

_elapsed___errors_____ops(total)___ops/sec(cum)__avg(ms)__p50(ms)__p95(ms)__p99(ms)_pMax(ms)__result
  600.0s        0         492516          820.9    116.9     67.1    419.4    570.4   1006.6

`DefaultVectorizeRowCountThreshold = 1000`

_elapsed___errors_____ops(total)___ops/sec(cum)__avg(ms)__p50(ms)__p95(ms)__p99(ms)_pMax(ms)__total
  600.0s        0          56267           93.8     17.8      8.4     67.1    109.1    302.0  insert

_elapsed___errors_____ops(total)___ops/sec(cum)__avg(ms)__p50(ms)__p95(ms)__p99(ms)_pMax(ms)__total
  600.0s        0        1068114         1780.2     53.0     37.7    159.4    234.9    570.4  scan

_elapsed___errors_____ops(total)___ops/sec(cum)__avg(ms)__p50(ms)__p95(ms)__p99(ms)_pMax(ms)__result
  600.0s        0        1124381         1874.0     51.2     33.6    159.4    226.5    570.4

Reverting the setting from 0 to 1000 increases the ops/sec from 821 to 1874, in line with the YCSB/E graph drop on December 4th.

@yuzefovich @asubiotto You may want to have a look at this.

YCSB/E performance regression on December 4th #62198

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions