Skippable exception applied to whole chunk and skip limit not strictly honoured for transactional reader [BATCH-1875]

[spring-projects-issues]

hari iyer opened BATCH-1875 and commented

When the reader is non-transactional, and the exception is skippable, Spring Batch does a good job of isolating the failing item(s) in a chunk write by rescoping the transaction around each item so that only the failing item is skipped. This does not seem to be the behavior when the reader is transactional. When I use JmsItemReader and one of the items in a chunk fails with a skippable exception during write, all items in the chunk get skipped. Is this the expected behavior?

---

Affects: 2.1.8

[spring-projects-issues]

Dave Syer commented

That was intentional I think. For a transactional reader you don't have much control (or any in general) over the order in which failed chunks are re-read, so the scanning approach we use in the non-transactional case doesn't have a well-defined outcome. It's not perfect, for sure, but I think transactional readers are not very common in the wild, so if you have one and a nice use case for testing, feel free to suggest changes.

[spring-projects-issues]

hari iyer commented

There are 2 use-cases that I am implementing.

1. Remote chunking using the spring-batch-integration components
2. Asynchronous item processing with chunking; what you called the BatchMessageListenerContainer.
   I created my own version of the listener message chunking for which I took inspiration from the remote chunking implementation. I extended the DefaultMessageListenerContainer and had it group messages into a single Message (a Message implementation that is just a container for a list of messages) and invoke a listener that uses the FaultTolerantChunkProcessor (configured on the same lines as the RemoteChunkHandlerFactoryBean)
   I tested this and found the reported behavior, so I tested with a JmsItemReader to make sure it was not something I did.
   I presume the remote chunking slave would behave the same way?
   Right now, it seems like the only way to use these and yet avoid this behavior is to do everything in the processor (where the skipping does work) and use a no-op writer to satisfy the step configuration.
   What do you suggest?