]

Under burst CI load the plugin registry returns HTTP 429 with a
Retry-After hint that the client previously ignored, so every attempt
landed inside the same rate-limit window and Nextflow startup aborted.

Two defensive layers around the registry metadata prefetch in
nf-commons, paired with a dependency bump that brings Retry-After
honoring into the HTTP layer:

  1. PluginUpdater.prefetchMetadata filters out plugins that are
     pinned to a specific version and already installed locally. When
     every requested plugin matches, the registry call is skipped
     entirely - no network round-trip on every Nextflow invocation in
     the common CI case (pinned versions, plugins already cached).
     Unpinned specs still go through so latest-release resolution
     keeps working.

  2. HttpPluginRepository.prefetch wraps the fetch in a try/catch on
     PluginRuntimeException, logs a warn, and leaves an empty plugins
     map. Prefetch is an optimization, not a hard requirement: a
     sustained registry rate-limit window or true outage must not
     abort Nextflow startup. The lazy getPlugin(id) path still
     surfaces per-plugin errors when an actually-missing plugin is
     needed downstream. The plugins field is initialised eagerly so
     getPlugin and refresh are safe to call when prefetch was skipped
     or failed.

To get HxClient to honor Retry-After on 429/503 the build picks up
[email protected] and [email protected], which require Failsafe >=3.2.1
(the new RetryPolicyBuilder.handleIf takes CheckedPredicate, and
ExecutionAttemptedEvent renamed getLastFailure to getLastException).
Bumping Failsafe 3.1.0 -> 3.3.2 surfaces a small amount of API drift
in pre-existing call sites that is migrated mechanically:

  - event.lastFailure -> event.lastException
  - Predicate<? extends Throwable> -> CheckedPredicate<? extends Throwable>
    on RetryPolicyBuilder.handleIf

Sites updated: PluginUpdater, PublishDir, GridTaskHandler, K8sClient,
BatchClient (nf-google), AzBatchService, AzFileSystem (nf-azure), plus
the matching AzBatchServiceTest Mock declaration.

Signed-off-by: Paolo Di Tommaso <[email protected]>

Defence-in-depth guard at PluginsFacade.start(): when NXF_OFFLINE=true
the registry call must never happen. The current code achieves this
transitively (PluginUpdater wraps repos with LocalUpdateRepository in
offline mode, so prefetchMetadata's PrefetchUpdateRepository iteration
is a no-op), but an explicit if(!offline) at the call site makes the
no-network guarantee unmissable in the code and protects against
regressions if a remote PrefetchUpdateRepository is later wired in
offline mode by mistake.

Signed-off-by: Paolo Di Tommaso <[email protected]>

Address review feedback: use an early-return guard instead of nesting
the try/catch inside the (plugins not empty) check. In Groovy an empty
or null list is falsy, so `!plugins` covers both cases.

Signed-off-by: Paolo Di Tommaso <[email protected]>

The skip-if-installed optimization in `prefetchMetadata` relied on
`pluginManager.getPlugin()`, but at prefetch time the local plugins have
not been loaded into the manager yet (its per-run root is still empty),
so the check always returned false and the `/v1/plugins/dependencies`
round-trip happened on every startup even when the plugin was cached.

Check the on-disk plugin store instead, mirroring `installPlugin()` which
reuses the cached copy whenever the `<id>-<version>` store directory
exists (it only downloads when missing). Its presence is the authoritative
signal that no remote metadata is required to start the plugin.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
Signed-off-by: jorgee <[email protected]>