sql: regression on EndToEnd/ored-preds

[yuzefovich]

Consider the following comparison between 22.1(edc5865) and 22.2(385bf90) of BenchmarkEndToEnd from pkg/sql/opt/bench:

EndToEnd/ored-preds-100/Simple-24                   2.53ms ± 1%    2.86ms ± 1%  +13.11%  (p=0.000 n=10+10)
EndToEnd/ored-preds-100/Prepared-24                 1.33ms ± 1%    1.47ms ± 1%  +10.92%  (p=0.000 n=10+8)
EndToEnd/ored-preds-using-params-100/Simple-24      4.35ms ± 1%    4.92ms ± 1%  +12.98%  (p=0.000 n=9+10)
EndToEnd/ored-preds-using-params-100/Prepared-24    2.85ms ± 1%    3.15ms ± 1%  +10.62%  (p=0.000 n=10+10)

(Full comparison is available here. The comparison was made with the following diff on 22.1 SHA:

--- a/pkg/util/goschedstats/runtime_go1.17.go
+++ b/pkg/util/goschedstats/runtime_go1.17.go
@@ -12,8 +12,8 @@
 // and go1.18. Before allowing newer versions, please check that the structures
 // still match with those in go/src/runtime.
 
-//go:build gc && go1.17 && !go1.19
-// +build gc,go1.17,!go1.19
+//go:build gc && go1.17 && !go1.20
+// +build gc,go1.17,!go1.20
 
 package goschedstats

to allow usage of go 1.19. The comparison was made with ./scripts/bench that used make under the hood with -benchtime=3s.)

Informs #87685. The regression doesn't seem that severe, so not marking it as a GA-blocker.

Jira issue: CRDB-20298

[mgartner]

We should try to get to the bottom of this, but I think this is a pretty minor regression with a very specific edge-case, so it shouldn't block the release.

[mgartner]

The regressions are more mild for me when I run on a GCE worker:

name                                              old time/op  new time/op  delta
EndToEnd/ored-preds-100/Simple-24                 2.46ms ± 0%  2.65ms ± 1%  +7.69%  (p=0.016 n=4+5)
EndToEnd/ored-preds-100/Prepared-24               1.22ms ± 1%  1.29ms ± 0%  +5.71%  (p=0.008 n=5+5)
EndToEnd/ored-preds-using-params-100/Simple-24    4.16ms ± 1%  4.49ms ± 2%  +7.92%  (p=0.008 n=5+5)
EndToEnd/ored-preds-using-params-100/Prepared-24  2.77ms ± 0%  2.89ms ± 0%  +4.51%  (p=0.008 n=5+5)

I'l try to bisect to a commit now.

[mgartner]

I was unable to easily bisect this. On many of the commits between the two commits listed in the description, I'm unable to run benchmarks - make spins forever.

[yuzefovich]

I also ran it on the gceworker with go 1.19.1.

[mgartner]

I'm going to close this. This regression seems very minor and the regressed benchmark is very contrived. We've already made significant improvements to this benchmark in the past, which dwarf this regression.