You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 36ad681
Browse filesBrowse the repository at this point in the historyBrowse files
lane-u5a-2359
committed
test: release the lock and reap the writer when a merge row fails
The waiting rows in this block hold the lock in this process and launch a writer
that is refused until they let go, and every release and every reap in them sits
inline after an assertion, so an assertion that lands between the acquire and the
release leaves this process holding the lock and the writer it launched still
contending for it. That is the half of the reply on row 4544685126 that the
longer retry window does not answer: more attempts change how long a writer is
willing to wait, not what becomes of the lock and the child when a row stops
early.
Guard the failure path for the block those rows live in, in the shape the row
named for letting go of both locks and the renewal timer already uses there,
releasing only what has not been released yet and then awaiting the writer, which
is admitted the moment the artefact is gone and exits on its own.
Pin it with two rows, because a guard nobody has failed into is a shape and not a
result. The first forces the failure these rows can hit, an assertion that lands
while this process still holds the lock and the writer it launched is still
refused, and records that at that moment the artefact is still there and the
writer has no outcome line, which is the hole and not the fix. The second runs
after the guard and asserts what the guard left, namely no artefact, a writer
that was admitted, wrote, released and exited, and the same path acquirable
again. Without the guard the second row fails and the first passes; with it both
pass and the file suite carries 31 passed and 8 skipped of 39 where it carried 29
passed and 8 skipped of 37.
No production file changed, so the priced unit delta measured against the unit
base is unchanged, and the mutation leg for these commits is still unmeasured
locally.
0 commit comments