Misses
gstack 1.91.138.0: the cheap PR screen lets fewer bad changes through than planned
screen-provisionalgstack now replays its real pull request history. Its cheap screening rule lets a bad change through on under 3 percent of pull requests, under half the 6 percent its plan assumed. The rule is still provisional.
Trains
Merge trains leave half full, so the gate runs more often
Merge trains fill to about two pull requests, not five. So the gate runs 22 to 29 times a week, well above the 12 to 20 the train model assumed.
Waits
With trains, a ready PR waits longer to merge
Trains also cost time. Today, the slowest tenth of ready pull requests merge within about three hours. Through trains, that stretches to five to eight hours.
Reds
Red builds are now sorted by code: caused by the PR, or from outside
Failed builds are now sorted in code: caused by the pull request, or from outside. It matches all 22 from the earlier study. Thirteen of 92 merged pull requests broke the build themselves, as the plan expected.
Your repo
Every gate still runs what it ran before
CHANGELOG.mdgstack 1.91.138.0 release notesNone of this changes what any gate runs. How long does a ready pull request wait to merge in your repo?
