gstack 1.91.137.0 measures what its paid PR gate costs, for free

1 hour ago

gstack 1.91.137.0 adds a free replay that prices gstack's paid PR gate from recorded runs (about $93 a full gate, almost all of it Fable 5.1), compares the cost per merged PR under three test-selection policies, and adds a seeded merge-train cost model. Every gate still runs what it ran before. Release notes: https://github.com/garrytan/gstack/blob/main/CHANGELOG.md#1911370---2026-10-18 How to run it: https://github.com/garrytan/gstack/blob/main/docs/TESTING_INTERNALS.md#replay-gate-cost

Ask

Ask about this presentation

Answers are generated from this presentation.

Chapters

  1. 0:00Gate cost
  2. 0:13Per merge
  3. 0:28Merge train
  4. 0:43Your gate
Show transcript

Gate cost

gstack logo

gstack now prices its paid PR gate for free, from runs it already recorded

$93.14
recorded per full gate
97.1%
spent on Fable 5.1, the model under test
Gate costPer mergeMerge trainYour gate
garrytan/gstack · CHANGELOG.md · Oct 18, 2026

gstack now prices its paid pull request gate for free, from runs it already recorded. A full gate costs about 93 dollars, almost all of it spent on Fable 5.1, the model under test.

Per merge

A screening rule would cut the gate bill per merged PR almost in half

Today's selector
$78.67
Exercised-map shadow
$64.27
Screening rulescreen-provisional
$45.17
Average per merged PR, 92 squash merges, Oct 3 to 18
Gate costPer mergeMerge trainYour gate
garrytan/gstack · CHANGELOG.md · Oct 18, 2026

Across 92 merges, today's test selection costs about 79 dollars per merged pull request. A screening rule would cut that to about 45, though those figures stay provisional until a real hunk classifier lands.

Merge train

A seeded merge-train model prices screening by the week

Defects in 6% of screened PRs, 25% unscreened
Gate flakes at 10% or 15%
Each red build pays a $6 bisect, and the culprit is requeued
Gate costPer mergeMerge trainYour gate
garrytan/gstack · CHANGELOG.md · Oct 18, 2026

A second model simulates a merge train, where unscreened pull requests carry defects four times as often and gates flake. Each red build pays for a bisect. It prices screening all, half, or none.

Your gate

Every gate still runs what it ran before

CHANGELOG.mdfirst step of a plan to cut the paid gate bill about 4x
What does one full run of your own test suite cost?
Gate costPer mergeMerge trainYour gate
garrytan/gstack · CHANGELOG.md · Oct 18, 2026

Every gate still runs what it ran before. This is the first step of a plan to cut the paid gate bill to about a quarter. What does one full run of your own test suite cost?