amadan
rollout sim: the rewake always wakes another instance when the marked build has no record (M17 on three seeds at 300)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2 files changed,
+5
−5
internal/rollout/sim_test.go+3 −3internal/rollout/testdata/rolloutsim/mutate.sh+2 −2
diff --git a/internal/rollout/sim_test.go b/internal/rollout/sim_test.go| index f55457041..79b18957e 100644 |
| --- a/internal/rollout/sim_test.go |
| +++ b/internal/rollout/sim_test.go |
| @@ -1884,8 +1884,8 @@ func (s *sim) retryHeld() { |
| // before another restriction of its app moves the generation; here the two |
| // halves are one step apart. |
| // |
| -// When X has no failure record yet, half the time the wake is instead |
| -// another asleep instance's on the same channel, one that never served X, |
| +// When X has no failure record yet, the wake is instead another asleep |
| +// instance's on the same channel, one that never served X, when there is one, |
| // and its first wake of X, if it runs, fails its health check (rewakeFail, |
| // as a flaky build may: a build with a marker and no record has seldom |
| // failed anywhere). That failure writes X's first record, which holds every |
| @@ -1917,7 +1917,7 @@ func (s *sim) rewakeArm() { |
| o = append(o, h2) |
| } |
| } |
| - if len(o) > 0 && s.gr.Intn(2) == 0 { |
| + if len(o) > 0 { |
| s.rewake, s.rewakeFail = o[s.gr.Intn(len(o))].name, x |
| } |
| } |
diff --git a/internal/rollout/testdata/rolloutsim/mutate.sh b/internal/rollout/testdata/rolloutsim/mutate.sh| index 0e1b360b8..4f7dd526e 100755 |
| --- a/internal/rollout/testdata/rolloutsim/mutate.sh |
| +++ b/internal/rollout/testdata/rolloutsim/mutate.sh |
| @@ -137,8 +137,8 @@ TIMEOUT_MIN=$((SEEDS * 3 / 100 + 20)) |
| # new record as a fact of every host. The random schedule seldom |
| # lines this up (a build with a marker and no record has seldom |
| # failed anywhere), so the simulation's rewakeArm makes it: when X |
| -# has no record, the wake after the Retry is sometimes another |
| -# instance's first wake of X, made to fail. Like M16 it rests on |
| +# has no record, the wake after the Retry is another instance's |
| +# first wake of X, made to fail. Like M16 it rests on |
| # few seeds at 300, and the same rule holds: a MISSED at 300 seeds |
| # or fewer is re-run at 3000 before it is read as a regression. |
| # (Removing the touch outright leaves the generation zero when the |