Revert lib-data-stream-redis to 1.5.0 and lib-cmd-queue-redis to 0.4.0

[pditommaso]

Restores both module trees to their state at 8bad0f5 ([release] lib-data-stream-redis@1.5.0, 14 May 2026) — the commit where stream was 1.5.0 and cmd-queue 0.4.0, so a single revert point covers both.

No [release] marker: this is not meant to republish. publish.sh skips versions already in the repo, so 1.5.0 / 0.4.0 would be no-ops anyway. Every source and test file in both modules is byte-identical to 8bad0f5 — only changelog.txt and build.gradle differ.

What is undone

PR	commit	undone
#84	bdf9374	async, non-blocking consumer processing with heartbeat lease + poll/renew/ack/release SPI → stream 2.0.0, cmd-queue 0.5.0 (plus doc follow-ups 11fdd3a, 86ae9ac)
#87	6c7b171	retry command on handler exception instead of terminal-failing it → 0.5.1
#89	e54229e	error tracking on CommandState (errorsCount, error, modifiedAt) → 0.6.0 (plus 1f124f0)
#86	f3f4ac4	migration to lib-data-workqueue(-redis) and CommandStatus SUBMITTED→PENDING / RUNNING→PROCESSING → 0.7.0
#91	e8fe916	README terminology
#94	d38afb0	only the parts that touched these two modules' sources

lib-cmd-queue-redis therefore depends on lib-data-stream-redis again.

Scope decisions

• lib-data-workqueue / lib-data-workqueue-redis are left in place, untouched. They carry the lease-based design forward and were never published (the release list in .github/workflows/build.yml does not include them). They are now unwired from cmd-queue but self-contained — separate packages (io.seqera.data.workqueue.* vs io.seqera.data.stream.*), so no clash.

• Two build-infra bits from build(java): Java 25 for Micronaut libs, Java 17 elsewhere; drop virtual-thread default; fix api leak #94 are deliberately kept rather than reverted, since they are repo-wide conventions and not module API:

  • the io.seqera.micronaut-library-conventions plugin id (Java 25 target for Micronaut modules)
  • Groovy 4.0.31 for stream's test dependencies — 4.0.24 cannot run on a JDK 25 toolchain

  Everything else in the two trees is byte-identical to 8bad0f5.

History preservation

• Each changelog keeps a REVERTED entry naming the withdrawn versions and their PRs, instead of silently rewinding.
• Pre-revert state is pinned on archive/workqueue-pre-revert (2ecc744), which also carries the unmerged invocation-lease rework.
• The parked workqueue modules carry a not-published banner instead of their stale "split of lib-data-stream-redis 1.6.0" callout (a version this PR removes), and the migration spec is marked HISTORICAL.

Fixes that must be re-landed (not carried by this PR)

This is a pure revert, so two fixes from the reverted range are dropped. Both are separable from the async/workqueue work that was rejected, and both belong in #101's 0.4.1 — sched's fix/ordered-shutdown-drain (f8f09293f, sched#888) already pins 0.4.1, so they arrive with no extra version bump:

1. [release] Retry command on handler exception instead of terminal-failing it #87 (6c7b171) — a thrown handler is terminal again here (store.save(state.failed(...)); return true), which strands work when a transient error hits the handler (sched#712). Separable: RedisMessageStream.consume only xack/xdels on true, so return false parks the entry in the PEL for XAUTOCLAIM, and the catch never persists started(), so status stays SUBMITTED and the next delivery re-enters execute(). Drain in-flight work before releasing the queue #101's 0.4.1 currently carries the 0.4.0 catch verbatim, so its drain narrows the window without closing it.
2. build(java): Java 25 for Micronaut libs, Java 17 elsewhere; drop virtual-thread default; fix api leak #94's api-scope fix — CommandQueue publicly extends AbstractMessageStream and exposes MessageConsumer, but the revert restores implementation project(':lib-data-stream-redis'). Same leak build(java): Java 25 for Micronaut libs, Java 17 elsewhere; drop virtual-thread default; fix api leak #94 fixed for the workqueue dep; breaks downstream subclassers in Gradle composite builds.

Audited for anything else: #87 is the only fix-shaped entry across the whole reverted changelog range, #86's cmd-queue diff is pure rename/migration, and the 1.2.0 xautoclaim starvation fix is intact in the reverted tree.

Downgrade note (recorded in the cmd-queue changelog)

Not a rollout prerequisite. 0.7.0 was never deployed — sched pinned 0.5.1 → 0.6.0 → 0.4.0 (a39502274, sched#772) and no other repo depends on this module. The downgrade that did happen, 0.6.0 → 0.4.0, is safe: JacksonEncodingStrategy disables FAIL_ON_UNKNOWN_PROPERTIES, so 0.6.0's errorsCount / error / modifiedAt are ignored on read.

For the 0.7.0 case only: it persists state with the renamed PENDING/PROCESSING values and a 7-day TTL. PROCESSING throws in store.findById() at step 1 of processCommand, outside the try/catch, so the entry is never acked — the listener survives but the entry re-claims and re-throws every claim-timeout, a poison entry in an ERROR loop rather than a silent skip. Self-heals at the TTL; flush the affected keys to downgrade sooner.

Verification

• ./gradlew assemble — BUILD SUCCESSFUL (all 134 tasks)
• :lib-data-stream-redis:test — 15 tests, 0 failures
• :lib-cmd-queue-redis:test — 15 tests, 0 failures
• workqueue modules still compile

🤖 Generated with Claude Code

[jordeu]

Not blocking — the revert is correct as it stands and should land. This is a follow-up request, not a change to this PR.

These three lines are what 6c7b171 (#87) replaced with log; return false, and that commit was written for sched#888's mechanism 14 days before the incident: Hikari pool closed during shutdown, handler throws, Redis state write still succeeds, command acked terminally, task stranded in PENDING forever (sched#712).

It is also separable from #84's async model, which is what sched#772 actually rejected. On 0.4.0/1.5.0 the false return means consume() skips xack/xdel, so the entry stays in the PEL for XAUTOCLAIM to hand to a live pod — sched#888's proposed primary outcome. And because this catch doesn't persist started() (unlike the timeout and explicit-RUNNING paths), status stays SUBMITTED and the next delivery re-enters execute(). That is the "re-run execute() later" primitive sched#889 §1 says 0.4.0 lacks — true of the return-running() path, not of the throw path with #87 applied.

So the ask is: land this, then cut a 0.4.1 = 0.4.0 + this one hunk from this same tree. sched then fixes #888 with a version bump. Java target checked: sched-app is on 25, so a republish from this tree loads fine.

Separately, #87 is worth naming in the risk section rather than only in the undone table — sched#772 re-accepted this regression on the basis that ReconcileCron would re-drive stranded tasks, and sched#889 shows 6d36bad0b never landed on master.

[jordeu]

Verified this has no live trigger, which is worth saying here so it doesn't read as a rollout prerequisite: sched's pin went 0.5.1 → 0.6.0 → 0.4.0 (a39502274), so 0.7.0 was never deployed, and no other repo pins this module (platform/wave use lib-data-stream-redis 1.1.1/1.1.2).

The direction that did happen — 0.6.0 → 0.4.0 on 22 Jul — is safe and worth recording: JacksonEncodingStrategy disables FAIL_ON_UNKNOWN_PROPERTIES, so 0.6.0's errorsCount / error / modifiedAt are ignored. Right now a reader can't tell whether the downgrade they already performed was OK.

Small precision on the symptom: PROCESSING throws in store.findById() at processCommand step 1, outside the try/catch. AbstractMessageStream.consumeOne catches Throwable, so the consumer survives, but the entry is never acked — it re-claims and re-throws every claim-timeout until the 7-day TTL drops the key. Poison entry with an ERROR loop rather than a silent skip; self-heals at TTL, so the advice itself is right.

[jordeu]

Confirmed self-contained — only settings.gradle:41-42 and lib-data-workqueue-redis/build.gradle:30 reference them, no cross-package imports.

Two stale bits now that nothing consumes them: VERSION is 1.0.0 and lib-data-workqueue-redis/README.md:20-21 tells readers to depend on io.seqera:lib-data-workqueue:1.0.0, which was never published; and that README plus docs/superpowers/specs/2026-07-11-workqueue-rename-migration.md describe these as the split of stream 1.6.0/2.0.0, versions this PR removes from the tree. A one-line "parked, not published" banner would save the next reader.

(Nit: the release list is in .github/workflows/build.yml, not publish.sh.)