Repository navigation
Grouped condition waits cannot replay after a physical reopen #278
Description
Found while checking #271 (part of #269) for the grouped case General raised from sdk-rust#188.
Problem
A condition wait inside all() / parallel() / a selection is re-sent with its parallel_group_path after a false wake, like an ungrouped wait. The Server records that physical reopen at the next sequence, outside the group's parallel_group_base_sequence / parallel_group_size bounds. On the next task, PHP replay refuses the history with NonDeterministicWorkflow: Parallel-group history contains invalid identity, bounds, or path fields. This happens on main too, with the full timeout, so any grouped timed or untimed condition wait that gets a false wake leaves the run stuck.
Probe (all([deferCondition(..., 'approval', 300), deferTimer(1000)])): opened at seq 1 (group base 1, size 2), a wrong signal, ConditionWaitSatisfied seq 1, then a reopen ConditionWaitOpened seq 3 with the same group path. Replay fails on main and on #271, whether the reopen records 300 or 280.
Scope
- Bind a grouped physical reopen to its original member (same group path, key, fingerprint, previous member satisfied or still open), as sdk-rust's
recorded_condition_group_pathdoes. Allow a recorded timeout equal to or smaller than the original, both present or both absent; a longer one is still refused. - Then apply Condition-wait reopen restarts its timeout and can leave the earlier wait open on native Server #269's remaining-time rule to grouped re-sends. Keep a condition wait's deadline across false wakes #271 deliberately leaves grouped waits on the authored timeout until replay is qualified.
- Tests and a corpus fixture: a grouped false wake reopens with the remaining time and the next task replays; a longer reopen is refused; selection winners are unchanged.
Check what native records for a grouped reopen (sequence and path) before implementing.
Activity
Keep grouped condition reopens on the authored timeout
Never link condition waits that belong to a group
rmcdaniel commentedon Nov 4, 2026
Reachability check before implementing (sdk-php, 2026-11-04), against published native Server 3.0.5 (
durableworkflow/server:3.0.5, SQLite, token auth) with the current main SDK (which sends occurrence ids):all([deferCondition(..., 300), deferTimer(600)])development_parallel_group_unqualifiedselect([deferCondition(..., 300), deferTimer(600)])development_selection_shape_unqualifiedselect([deferCondition(...), deferCondition(...)])development_selection_shape_unqualifiedNative
all_groups.rsqualifies only activity and timer members, and the selection shapes above are refused, so native never records a grouped condition wait from PHP and never satisfies or reopens one. Per sdk-rust's live check (General, lobby), PHP Server 2.11.8 records noConditionWaitSatisfiedfor its grouped layout, so it triggers no re-send either.So the failure in this issue is not reachable on any published runtime today: it needs a runtime that both accepts a grouped condition wait and satisfies it on a false wake. #271 already keeps grouped re-sends on the authored timeout. Proposed: keep this open, blocked until a runtime qualifies grouped condition-wait shapes, then implement the binding (via
condition_wait_occurrence_id, as sdk-rust does), using that runtime's recorded layout as the fixture.rmcdaniel commentedon Nov 4, 2026
sdk-php owns the PHP replay fix and source/corpus follow-through. The concrete missing input is the first supported grouped-condition reopen shape, including canonical bounds/path, occurrence binding and source-produced history. I asked Server on its owning contract issue: durable-workflow/server#325 (comment) . Server owns that admission/history contract; General owns its priority after the current P1 → timer cleanup → flat-child-all work. This grouped-condition defect remains separate from #185's child-only barrier.
Next action: reconcile the Server contract response, then implement the smallest PHP grouped-reopen fix with a fail-before/pass-after replay fixture, remaining-time and longer-timeout controls, and unchanged selection-winner behavior. I have checked the existing PHP failure and retained its sequence/path/bounds evidence; I have requested the producer contract rather than starting an unapproved native experiment or treating the hold as completion. No package/release or runtime experiment is claimed.
rmcdaniel commentedon Nov 5, 2026
I rechecked the retained PHP group bounds/path failure and requested the missing producer contract from the current Server owner under #325. No first supported grouped-condition reopen history was returned by this handoff; flat child-only #1026 source acceptance does not qualify grouped conditions.
Server owns canonical supported bounds/path/occurrence history and General its priority. sdk-php owns the smallest grouped-reopen replay correction and official fail-before/pass-after fixture once that history exists, including remaining-time, longer-timeout refusal and unchanged selection winners. No duplicate runtime experiment or invented producer layout is started.
2026-11-05 accepted-source and dependency reconciliation.
I completed the finite flat child-only consumer adapter through #312 actual-main acceptance. That adapter and Server's child/run/link/queue contract provide no grouped-condition reopen bounds/path/occurrence history, so they do not close this defect.
Next action / owner: Server returns the first supported canonical grouped-condition reopen history and General its source priority; sdk-php then implements the smallest replay correction with old-fail/new-pass fixtures, remaining-time and longer-timeout refusal, and unchanged selection winners. The current Server owner has been asked for that concrete retained source slot. No invented producer layout, duplicate runtime experiment or central-security waiting condition was added.
Current upstream owner return: Server explicitly owns the first versioned grouped-condition bounds/path/call/occurrence/history and physical wait/timer proposal after its active train. Current native source still refuses that shape and no new settled capability exists. Durable Server return.
rmcdaniel commentedon Nov 6, 2026
2026-11-06 owner reconciliation.
I rechecked Server's bounded grouped-condition source ownership. The newer native source checkpoint returns no canonical grouped reopen bounds/path/call/occurrence history or physical wait/timer proposal. The accepted flat-child adapter and current credential/docs source work do not establish that shape.
Server owns the next source action: return the first supported grouped-condition reopen history and versioned bounds/path/occurrence contract, with General setting its priority. sdk-php then owns the smallest replay correction and old-fail/new-pass corpus control, including remaining time, longer-timeout refusal and unchanged selection winners. I retained this exact request with Server/General. Current native grouped-condition refusals remain until the producer shape is concrete; no layout is invented and no duplicate installed/runtime run was started.
rmcdaniel commentedon Nov 6, 2026
I re-read the grouped physical-reopen failure and Server's retained bounds/path/occurrence ownership while reconciling the new P0 work. No accepted canonical native grouped-reopen producer history has been returned; source clocks and the flat-child adapter do not supply that missing representation. #335's stable internal performance cut leaves this replay path unchanged.
Next action / owners: Server returns the supported versioned grouped-condition reopen history plus physical wait/timer contract, with General ordering that source work. sdk-php owns the smallest bounds/member replay correction and immutable-base corpus negative, remaining-time and longer-timeout refusal, and unchanged selection winners once those actual producer bytes exist. I retained the concrete producer request with the current Server owner and General. No invented group layout, runtime probe or claim that clock-only source fixes the reopen was added.
rmcdaniel commentedon Nov 7, 2026
Post-release boundary is checked: actual SDK3.9.5 still has no newly returned canonical grouped-condition physical-reopen producer history. The helper/grant patch cannot supply that missing representation. Next owner/action: Server returns the versioned grouped-reopen history plus physical wait/timer contract under General's existing ordering; sdk-php then owns the minimal bounds/member replay fix and immutable-base negative, remaining-time/longer-timeout controls and unchanged selection winners. I retained the producer request; no group layout or runtime claim was invented.
rmcdaniel commentedon Nov 7, 2026
Handoff audit: no new accepted canonical grouped-condition physical-reopen producer history has been returned. The newly completed carried-poll/pipeline source cohort does not supply that missing representation. I re-read this producer boundary and retained the existing request.
Next owner/action: Server returns versioned grouped-reopen history plus the physical wait/timer contract under General’s ordering; sdk-php then owns the minimal bounds/member replay fix, immutable-base negative, remaining-time/longer-timeout controls and unchanged winners. Server/General have been asked. No group layout or runtime result was invented.
rmcdaniel commentedon Nov 7, 2026
Recheck 2026-11-07 19:40 UTC: unchanged. No published runtime accepts a grouped condition wait, so the failure is still unreachable, and Server has not returned a grouped-reopen history/bounds contract.
Next owner / action: Server (contract, under General's ordering); sdk-php then does the replay fix with old-fail/new-pass fixture. Asked Server again in chat today.
rmcdaniel commentedon Nov 11, 2026
sdk-php (2026-11-11): native Server main now accepts
conditionas a grouped/selection member kind (completion_rules.rs), so this failure may now be reachable on published native rather than PHP-only. Next action (sdk-php): after 3.13.0, reproduce against published native 3.4.0 withall([deferCondition(...), deferTimer(...)])and a false wake, and capture what native records for the reopen (sequence and group path). If it reproduces, fix replay in the SDK like sdk-rust'srecorded_condition_group_path, with an old-fail/new-pass fixture. No Server contract is needed for that. If native refuses grouped reopens, record that here and close.rmcdaniel commentedon Nov 11, 2026
sdk-php (2026-11-11), probe against published native 3.4.0 (SQLite) with SDK main: a workflow running
$context->all([$context->deferCondition($approved, 'approval', 300), $context->deferTimer(3)])never reaches the false-wake reopen. Its first completion is refused with 422development_parallel_group_unqualified(rust/src/runtime/all_groups.rs), and the managed worker exits. So the replay failure described here can't happen on published native. Grouped condition waits aren't accepted there at all.State: blocked on native accepting grouped condition waits. I've asked server in #sdk-php whether that's planned. Separate user-facing gap: the worker process dies on that 422 instead of failing just the task with a clear message. If native won't support it soon, the SDK should refuse this shape at authoring time with a clear error. Owner: server (decision), then sdk-php.
rmcdaniel commentedon Nov 11, 2026
General decision, 2026-11-11 (recorded by sdk-php): refuse a new condition or signal-wait member inside
all()/parallel()only when the connected Server is native (cluster info implementationrust) and doesn't advertise a grouped-condition capability. The refusal is a clear task failure: "condition waits inside all()/parallel() aren't supported; await the condition outside the group; server#1142; refused by Server ". PHP Server 2.x keeps today's behavior. Already-recorded groups keep replaying everywhere. It ships in 3.13.1 with tests for both Servers. Server is asked to name the capability on server#1142 so the refusal lifts automatically. Next: sdk-php implements it after the 3.13.0 release.Merge pull request #386 from durable-workflow/fix/278-refuse-grouped-…
rmcdaniel commentedon Nov 11, 2026
Installed check PASS (2026-11-11 ~19:15 UTC):
durable-workflow/sdk3.13.1 from Packagist against publisheddurableworkflow/server:3.4.0. A workflow withall([deferCondition(..., 'approval', 300), deferTimer(3)])now fails its workflow task withgrouped_condition_wait_unsupported: condition waits inside all()/parallel() aren't supported; await the condition outside the group (server#1142). Refused by Server native 3.4.0.The worker process stays running (staterunning), where 3.12/3.13.0 exited on the 422. Released in 3.13.1.Closing per General's decision: users get a clear, recoverable failure today. The original replay concern (a grouped false-wake reopen) becomes reachable only once native accepts grouped condition waits. I've asked that server#1142 include the reopen sequence/path contract, and an SDK replay fixture will land with it.
Opt condition waits into the Server's occurrence deadline when advert…