We now additionally guard the completion query by event and target identifier to also only apply to publications that have not been completed yet. This will allow databases to optimize the query plan to apply simple comparisons (the date being null) over complex comparisons (the event payload) to reduce the intermediate results to process further and thus improve performance.
We now properly wrap the null value returned as a result from the RabbitMQ event externalization into a CompletableFuture. Tightened test case to prevent regressions going forward.
The default EventPublicationRegistry previously erroneously called find*In*completePublications(), i.e. returned incomplete publications when it was supposed to look up the completed ones. To be able to do the latter we need to extend EventPublicationRepository and introduce a query execution for all completed publications.
As we'd normally only chance API in major versions, but the bugfix needing to go into a bugfix release, the new method is introduced as default method rejecting the execution in case a currently existing repository implementation does not implement it. This allows us to adapt the implementations we ship to support the bugfix, but will require other implementations to ship an adapted version. This means we can ship a release that's not breaking assuming one of the official store implementations is used.
We now explicitly handle CompletableFuture instances returned from transactional event listeners by registering the completion handlers only on success, the debug logging on failure and immediately return the decorated instance.
Previously, a failed CompletableFuture instance would still have the publication marked completed as it doesn't cause any exception being thrown and thus ultimately ending up in the code path that issues the completion.
We now propagate potentially occurring exceptions in CompletionRegisteringAdvisor to make sure the standard exception handling facilities kick in, both for synchronous and asynchronous listener invocations.
This implies that during republication we need to handle exceptions to make sure that failing synchronous listeners to not prevent the submission of subsequent publications. We currently log such exceptions into error.
