Can any Dataflow experts / Googlers enlighten me as to why this happens? I shimmed the FnHarness for a Python pipeline with a Java external transform, and it seems that the ProcessBundleHandler receives different process-bundle-descriptor-%d for the same processor. This leads to the system defeating the cachedReaders setup (and doesn't evict previous cachedReaders, as it remains alive in those readers from what I've seen).
(The story about the Solace source continues...) Thanks Alex
