Repository navigation
3.3.1
Durable Workflow Server 3.3.1.
| Git tag | 3.3.1 → 0bcadca1bbab1d262f43bff11b802e31af9124cb |
| Image index | sha256:775e83d5584cd49722ecde379efa8be3cc0a0ba935f7d31131c70bcddb2ef788 in durableworkflow/server and ghcr.io/durable-workflow/server |
| Release run | https://github.com/durable-workflow/server/actions/runs/42550966502 |
Durable Workflow Server 3.3.1 (native)
Server 3.3.1 is a patch release of the native (Rust) Server, published as
durableworkflow/server:3.3.1 and ghcr.io/durable-workflow/server:3.3.1.
The published Compose file defaults to it. Rolling tags move in the separate
promotion step after qualification.
Upgrade
There is no schema upgrade from 3.3.0. MySQL, MariaDB and PostgreSQL remain
at native schema 11; SQLite remains at schema 10. The worker protocol stays
at 1.20, the control plane at version 2, and the capabilities are unchanged.
Existing clients need no SDK release for these internal query reductions.
For upgrades from 3.2.x or earlier, follow the
3.3.0 upgrade and rollback instructions, including the
backup, explicit schema-upgrade and condition-wait rollback constraints.
This patch does not relax those constraints or repair a mixed condition-wait
chain.
Changes
PR #1099 reduces
database statements on workflow transitions while retaining the existing
answers, transaction boundaries and lease authority:
- Upsert a run summary instead of checking its existence before writing it.
- Read a workflow poll's cached binding and result together.
- Aggregate the projection census in one query.
- Read the authored sequence counters together, retaining integer types on
MariaDB and zero for missing rows. - Reuse the waiting-status result already read in the transition.
- Reuse fenced worker registration within the transaction.
- Keep run tags within the existing transaction context.
The paired poll-cache writer proposed in
issue #1063 is
deferred. Binding and result writes retain their separate expiry-time origins.
The other items in the performance report remain tracked individually.
No benchmark speedup or latency improvement is claimed for this release.
The qualification tooling also verifies the published PHP replay matrix
under a read-only root, dropped capabilities and no-new-privileges. A private,
byte-identical PHP executable copy avoids inheriting the tooling image's file
capability; Server and SDK product bytes are not patched for the experiment.
Known limitations
The 3.3.0 known limitations still apply.
Source/backend query-count controls and published-artifact compatibility are
separate evidence; passing them does not reproduce the reporter's benchmark.