Page MenuHomePhabricator
Paste P96536

more_like track_total_hits perf bench
ActivePublic

Authored by dcausse on Tue, Sep 29, 10:01 AM.
Project Tags
None
Referenced Files
F106403132: image.png
Mon, Oct 5, 10:03 AM
F105415735: more_like track_total_hits perf bench
Tue, Sep 29, 10:01 AM
Subscribers
None
=== size=4 ===
track_total_hits=4 took n= 1000 mean= 146.5 p50= 131.0 p90= 261.0 p99= 503.1 max= 794.0
track_total_hits=4 wall_ms n= 1000 mean= 150.3 p50= 133.9 p90= 264.2 p99= 506.6 max= 797.2
track_total_hits=true took n= 1000 mean= 252.4 p50= 213.0 p90= 453.0 p99= 792.1 max= 1235.0
track_total_hits=true wall_ms n= 1000 mean= 256.6 p50= 216.9 p90= 456.2 p99= 795.4 max= 1237.6
4 vs true (paired over 500 docs, ratio = true/4, >1 means 4 is faster)
took ratio: median=1.59 p10=0.97 p90=3.63
wall ratio: median=1.61 p10=0.97 p90=3.51
4 strictly faster (took) on 438/500 docs (87.6%)
docs with different top ids: 163/500
took ratio by exact hits.total:
<1k n= 6 median=0.90 p90=1.49
1k-10k n= 37 median=1.03 p90=1.44
10k-100k n= 118 median=1.19 p90=1.67
>100k n= 339 median=2.12 p90=4.21

Event Timeline

For posterity this is the outcome of deploying this optimization to all wikis:

image.png (3,145×664 px, 266 KB)

P50 went from 200ms to ~150ms
P95 are slightly more stable in codfw (rarely above 500ms), for eqiad there's a definitely a gain but the P95 remains highly variable depending on the time of the day.