Repository navigation

Pace replication retries on one jittered schedule and bound subscription setup - #800

Merged

kriszyp merged 25 commits into

mainfrom

fix/replication-uniform-backoff

Sep 29, 2026

Commits on Sep 2, 2026
- Impose a uniform backoff discipline on replication retries
  
  kriszyp
  and
  claude
  committed
  f214bd2
- fix: carry the enriched payload on the armed subscription setup
  
  kriszyp
  and
  claude
  committed
  d2bb0c1
- Address pre-push review: fail-closed deadline, equal jitter, owned recovery timers
  
  kriszyp
  and
  claude
  committed
  2c3902c
- Harden replication retry admission and jitter
  
  kriszyp
  and
  codex
  committed
  3407086
- Keep the #446 stagger intact and scope connect-side cancellation
  
  kriszyp
  and
  claude
  committed
  890c9ac
- Self-heal a rejected worker-readiness on the shared schedule
  
  kriszyp
  and
  claude
  committed
  4c0b1ec
- Address the pre-push review: deadline, sweep completeness, and timer ownership
  
  kriszyp
  and
  claude
  committed
  9a50aef
- Fix the review's documentation, floor, and log-rate nits
  
  kriszyp
  and
  claude
  committed
  23ba669
- Reset the setup backoff on connect instead of attributing the report
  
  kriszyp
  and
  claude
  committed
  fba0d26
- Give the stall kick its own claim, and make the docs match the code
  
  kriszyp
  and
  claude
  committed
  0ea8e54
- Hand the stall throttle back when a reconnect cancels the kick
  
  kriszyp
  and
  claude
  committed
  537835b
- Put shouldFireStallKick below the function its description sat above
  
  kriszyp
  and
  claude
  committed
  7f3431d
- Resolve the node URL from one map lookup
  
  kriszyp
  and
  claude
  committed
  a5b6aa0
Commits on Sep 7, 2026
- Honor current replication connection ownership
  
  kriszyp
  and
  codex
  committed
  ff115a3
- Preserve catchup state across retry pacing
  
  kriszyp
  and
  codex
  committed
  3074e39
- Keep delayed recovery attempts authoritative
  
  kriszyp
  and
  codex
  committed
  da6513f
Commits on Sep 28, 2026
- Scope stall generations to the primary connection and keep setups off the orchestrator
  
  committed
  e0a58ae
- Merge remote-tracking branch 'origin/main' into fix/replication-uniform-backoff
  
  kriszyp
  and
  claude
  committed
  da43922
- Describe the unified socket-open report in the recovery design note
  
  kriszyp
  and
  claude
  committed
  1ba7ad2
- Keep an escalated pair out of the sweep ladder, and re-send self-catchup to a replacement worker
  
  kriszyp
  and
  claude
  committed
  130355c
- Bound the worker-restart probes, sample the per-peer repair warn, and trim comments
  
  kriszyp
  and
  claude
  committed
  eddd73e
- Space a reassignment sweep by the setups it armed, not a chained floor
  
  kriszyp
  and
  claude
  committed
  4028c1f
- Space a reassignment sweep by absolute fire times
  
  kriszyp
  and
  claude
  committed
  a55d365
- Put the blob-send 503 re-read on the shared backoff and define the subscription-setup schedule once
  
  kriszyp
  and
  claude
  committed
  a3a3b28
- Type the reconnect backoff as optional and name the exhaustion test for what it asserts
  
  kriszyp
  and
  claude
  committed
  bacd6b6

Navigation Menu

Repository navigation

Pace replication retries on one jittered schedule and bound subscription setup - #800

Pace replication retries on one jittered schedule and bound subscription setup#800
kriszyp merged 25 commits into
mainfrom
fix/replication-uniform-backoff

Commits on Sep 2, 2026

Impose a uniform backoff discipline on replication retries

fix: carry the enriched payload on the armed subscription setup

Address pre-push review: fail-closed deadline, equal jitter, owned recovery timers

Harden replication retry admission and jitter

Keep the #446 stagger intact and scope connect-side cancellation

Self-heal a rejected worker-readiness on the shared schedule

Address the pre-push review: deadline, sweep completeness, and timer ownership

Fix the review's documentation, floor, and log-rate nits

Reset the setup backoff on connect instead of attributing the report

Give the stall kick its own claim, and make the docs match the code

Hand the stall throttle back when a reconnect cancels the kick

Put shouldFireStallKick below the function its description sat above

Resolve the node URL from one map lookup

Commits on Sep 7, 2026

Honor current replication connection ownership

Preserve catchup state across retry pacing

Keep delayed recovery attempts authoritative

Commits on Sep 28, 2026

Scope stall generations to the primary connection and keep setups off the orchestrator

Merge remote-tracking branch 'origin/main' into fix/replication-uniform-backoff

Describe the unified socket-open report in the recovery design note

Keep an escalated pair out of the sweep ladder, and re-send self-catchup to a replacement worker

Bound the worker-restart probes, sample the per-peer repair warn, and trim comments

Space a reassignment sweep by the setups it armed, not a chained floor

Space a reassignment sweep by absolute fire times

Put the blob-send 503 re-read on the shared backoff and define the subscription-setup schedule once

Type the reconnect backoff as optional and name the exhaustion test for what it asserts