Repository navigation
3.11.0
Durable Workflow Server 3.11.0.
| Git tag | 3.11.0 → da38c0ede9b15046fcab2bb0f25412c85a3e3404 |
| Image index | sha256:39d218fc5f251a3e7c4ff1dcda06d9cc53f66c034f0c6d0f67e8371fd0cc97d6 in durableworkflow/server and ghcr.io/durable-workflow/server |
| Release run | https://github.com/durable-workflow/server/actions/runs/43521139955 |
Durable Workflow Server 3.11.0 (native)
Server 3.11.0 reuses completed plain timers when redriving a run that failed at
an activity. The successor resumes the failed activity with the original timer
clock. Its immutable images are
durableworkflow/server:3.11.0 and ghcr.io/durable-workflow/server:3.11.0;
the published Compose default selects 3.11.0. Rolling aliases move separately
after the published producer and applicable SDK replay qualification.
Fired timer reuse
A plain timer before the failed activity must have exactly one ordered
TimerScheduled / TimerFired pair at a unique step, with matching source timer
identity. The successor records the pair under one new shared timer ID. It
creates no timer row or timer task and does not wait for the timer again.
The reused TimerFired records the immediate predecessor in
reused_from_run_id. Its reused_recorded_at preserves the timer's original
execution time, including after another redrive. The original source history
remains unchanged. Earlier activity and marker steps keep their existing
reuse rules, and only the failed activity is resumed.
Redrive replay timestamps now retain nonzero fractional precision for reused
activities, markers and fired timers. The initial replay anchor retains the
persisted source run start with full precision. Repeated redrives preserve
already-recorded reuse timestamps as recorded. This change cannot reconstruct
fractions omitted by an earlier producer. Deadline formatting is unchanged.
Scope, group, cancellation-cleanup, condition and signal-timeout timers remain
outside this plain-timer slice and refuse as unsupported_timer. Incomplete,
reversed, duplicate or mismatched timer pairs refuse as incomplete_timer_step.
Step collisions, gaps and invalid failure boundaries retain their explicit
refusals. Cancelled timers and other unsupported history remain refused.
Describe and redrive use the same eligibility decision.
Upgrade and compatibility
From 3.10.x, no schema step is needed. The physical schema remains SQLite 14,
PostgreSQL and MariaDB 15, and MySQL 16. Follow the existing backup and upgrade
guidance, pin the new image, and verify readiness, reported version and worker
registration. Update API replicas consistently before relying on the expanded
redrive eligibility; older replicas retain their earlier refusal rules.
From 3.9.x or earlier, apply the applicable explicit schema steps in the
3.10 upgrade procedure.
This release does not convert a PHP database at startup. The existing takeover
runbook retains its exact historical version, backend and state limits.
Worker protocol 1.20, control-plane version 2, codec and command shapes are
unchanged. Existing command-key, completion-byte, storage-admission and
unsupported service-upsert boundaries remain in force. This release does not
add redrive support for signals, updates, child workflows or scoped/group timers.
Qualification and retained limits
The actual PHP, Python and Rust Worker/cold-process timer-clock proof is tracked
on #980. Source controls
alone do not establish those installed consumer outcomes. Producer source,
backend/corpus, immutable publication, Compose, scan/provenance and exact
published consumer evidence remain separately attributable. Core precision
correction and reference-corpus follow-through retain their separate record in
Workflow #992.
The MariaDB registration-timeout investigation
#1204 remains separate.
Cloud adoption, physical mapping and broader cancellation-policy qualification
retain their existing limits. PHP Server 2.11.8 remains retired and unpatched; see the
retirement/security notice.