MOVING handoff: hold new work back while draining, so the recycle happens within a round trip instead of at the deadline by mgravell · Pull Request #3335 · StackExchange/StackExchange.Redis

Handoff drain: cover the pinned successor, cap at half the remaining window, exact retired marker

mgravell

committed

04c6761

Handoff drain: held work survives the reconnect, per-bridge deadlines, cover ends on establishment, teardown outside the lock

mgravell

committed

b94531f

Handoff drain: cover ends when both bridges are up, no writer touches a retired connection, FailFast exemption scoped, recycle path retires what is current

mgravell

committed

0fc57a4

Handoff drain: one lifecycle for the FailFast shield and the timeout cover

mgravell

committed

77e01b8

Handoff drain: drop the front queue, lower protection on an empty bridge, leave a replacement that landed elsewhere alone

mgravell

committed

a5d4804

Handoff drain: guarded landed-on read, null-at-begin through the same path, give-up keeps the shield for diverted work

mgravell

committed

cef2b90

Handoff drain: the shield keys on durable facts about the connection, not on where the handoff's flow stands

mgravell

committed

95d507d

Handoff drain: protection stands down only for a bona fide replacement; one tick-deadline type

mgravell

committed

5b10b77

Handoff drain: a resolved retire makes the next connection the replacement; a hold that diverted nothing protects nothing

mgravell

committed

1924d9f

Handoff drain: target status dominates the replacement latch, the landed-on test applies after the retire, a lockless retire as the last resort

mgravell

committed

c58abb9

Handoff drain: a temporal fallback where endpoint identity is meaningless; the cover applies to held commands only

mgravell

committed

ee6ae39

Handoff drain: three-way classification with unknown deferring, the cover as a property of held commands, trust validated at use

mgravell

committed

2179c10

Handoff: one explicit phase per handoff in place of the parallel protection flags

mgravell

committed

5451bf9

Handoff: a periodic re-judgement, doomed dominating the pin, and the drain's bounds

mgravell

committed

54d49e5

Handoff: per-handoff deadlines, a second chance for the retire, and the last two wrap hazards

mgravell

committed

60da7a6

Handoff: edges - same-address successors, sweeps past covered heads, retire and stamp races, slicing only while a window is open

mgravell

committed

99a9ab8

Handoff: the claim is released once, the pre-hold queue is covered, and judgement needs evidence

mgravell

committed

c48c8e6

Handoff: doomed is doomed in every phase, and "a handoff exists" is the phase, not the drain's claim

mgravell

committed

ec4f853

Handoff: the cover follows the write time into transactions, and the pin is one generation-checked state

mgravell

committed

21ed775

Navigation Menu

Repository navigation

MOVING handoff: hold new work back while draining, so the recycle happens within a round trip instead of at the deadline - #3335

MOVING handoff: hold new work back while draining, so the recycle happens within a round trip instead of at the deadline#3335
mgravell wants to merge 27 commits into
mainfrom
marc/moving-handoff-quiesce

Commits on Nov 4, 2026

MOVING handoff: hold new work back while draining, so the recycle happens within a round trip instead of at the deadline

Handoff drain: cap the wait, recycle under the write lock, hold only live caller work

Handoff drain: retire only the connection the hold began on, under the lock, with a bounded wait; requeue raced writers at the head

Handoff drain: head inserts under the backlog lock, marker cleared only at the next hold, per-bridge release, cap floor

Handoff drain: survive the old connection dying mid-hold, cover held work's timeout, budget from what is left

Commits on Nov 5, 2026

Handoff drain: cover the pinned successor, cap at half the remaining window, exact retired marker

Handoff drain: held work survives the reconnect, per-bridge deadlines, cover ends on establishment, teardown outside the lock

Handoff drain: cover ends when both bridges are up, no writer touches a retired connection, FailFast exemption scoped, recycle path retires what is current

Handoff drain: one lifecycle for the FailFast shield and the timeout cover

Handoff drain: drop the front queue, lower protection on an empty bridge, leave a replacement that landed elsewhere alone

Handoff drain: guarded landed-on read, null-at-begin through the same path, give-up keeps the shield for diverted work

Handoff drain: the shield keys on durable facts about the connection, not on where the handoff's flow stands

Handoff drain: protection stands down only for a bona fide replacement; one tick-deadline type

Handoff drain: a resolved retire makes the next connection the replacement; a hold that diverted nothing protects nothing

Handoff drain: target status dominates the replacement latch, the landed-on test applies after the retire, a lockless retire as the last resort

Handoff drain: a temporal fallback where endpoint identity is meaningless; the cover applies to held commands only

Handoff drain: three-way classification with unknown deferring, the cover as a property of held commands, trust validated at use

Handoff: one explicit phase per handoff in place of the parallel protection flags

Handoff: a periodic re-judgement, doomed dominating the pin, and the drain's bounds

Handoff: per-handoff deadlines, a second chance for the retire, and the last two wrap hazards

Handoff: edges - same-address successors, sweeps past covered heads, retire and stamp races, slicing only while a window is open

Handoff: the claim is released once, the pre-hold queue is covered, and judgement needs evidence

Handoff: doomed is doomed in every phase, and "a handoff exists" is the phase, not the drain's claim

Handoff: the cover follows the write time into transactions, and the pin is one generation-checked state

Commits on Nov 6, 2026

Handoff: the replacement is whatever establishes once the target is gone; the contract; a two-MOVINGs matrix

Handoff: the post-enqueue re-stamp no longer depends on the first read finding no handoff

Tests: the two-MOVINGs matrix keeps its workers throttled in the drain cases