allthingssecurity commented on PR #29893:
URL: https://github.com/apache/camel/pull/29893#issuecomment-6875852963

   Thanks @davsclaus, all four points are in 2fa913579fa5.
   
   **[Important] handover after a stop.** Taken your first option. The informer 
now has `.exceptionHandler((started, t) -> started)`, so once it is started 
fabric8 retries any error by itself and lists again after a watch error, and 
the informer is stopped only when the task stops. I checked it against the 
8.0.0 jars with javap: `SharedIndexInformer.exceptionHandler(ExceptionHandler)` 
with `boolean retryAfterException(boolean isStarted, Throwable t)`; 
`Reflector.onException` passes "start future done and not failed" as 
`isStarted`, and on true calls `reconnect()` (exponential backoff), else 
completes the start and stop futures exceptionally. The default handler is 
`isStarted && !(t instanceof WatcherException)`, which is why a watch error 
other than 410 stopped it. The loop now only starts a new informer when one 
fails to start, with the 1 s to 30 s backoff, and each informer gets its own 
store, so nothing can hide in the store of a previous one.
   Test: 
`KubernetesInformerLoopTest.testUpdateQueuedWhileTheHandlerIsBusyIsReportedAfterAWatchError`.
 The handler blocks on the first update (as a reload would), a second update 
arrives and is queued, then the watch gets an event that is not a resource (a 
watch error other than 410), and the test waits until the client has closed 
that watch before it releases the handler. The queued update must still come, 
and the busy handler must not be interrupted. One thing I saw while writing it: 
fabric8 runs the re-list on the same serial executor as the handler, so after a 
watch error the informer lists again only once the reload in progress is done. 
Without the exception handler it fails (`the queued update of my-b was dropped 
==> expected: <my-b> but was: <null>`, and the earlier restart test fails too). 
The earlier restart test is now 
`testUpdateMadeBeforeAWatchErrorOtherThanGoneIsReported`: the informer lists 
again by itself.
   
   **[Suggestion] what the store keeps.** Each informer's store is a 
`ReducedStateItemStore` (in 8.0.0, with `NAME_KEY_STATE`), subclassed to skip 
names that do not match. It keeps only the name, namespace and resource 
version, which is what `ProcessorStore` compares, so the Secret data stays out 
of the long-lived cache. `onUpdate` still gets the full new object from the 
event, used only for its name.
   
   **[Suggestion] upgrade guide.** Added: the `list` permission does not widen 
what the ServiceAccount can read, as the old watch had no resource version and 
already got every existing Secret (or ConfigMap) as an ADDED event.
   
   **[Nit] stop test.** `testStopWhileTheInformerStartsEndsTheLoop` now asserts 
no watch with Awaitility `during(4 s)`, past the 3 s delay of the list.
   
   Tests: `KubernetesInformerLoopTest` 7, 
`SecretsReloadTriggerTaskWatchClosedTest` 5, 
`ConfigmapsReloadTriggerTaskWatchClosedTest` 5, all pass; without the exception 
handler 2 of the 7 loop tests fail. The PR description is updated to match.
   
   _Claude Code on behalf of allthingssecurity_
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to