ComposedSegmentedLoadWriteStore always parallelizes iterations, even when the consumer is a single thread doing a blocking iteration like cache.keySet().forEach(...). This will use lots of blocking threads, and will crowd out single-key cache operations, which are more sensitive to latency (because there are usually multiple single-key operations for each user request).

It would be interesting to extend the persistence SPI so that the store knows when the user requested parallel iteration, but in the meantime it's safer to iterate over the segments in sequence.