Cached state stores may forward records during a flush call, so Streams should flush the stores in topological order. Otherwise, Streams may flush a downstream store before an upstream one, resulting in sink results being committed without the corresponding state changelog updates being committed.
This behavior is partly responsible for the bug reported in
The fix is simply to flush the stores in topological order, then when the upstream store forwards records to a downstream stateful processor, the corresponding state changes will be correctly flushed as well.
An alternative would be to repeatedly call flush on all state stores until they report there is nothing left to flush, but this requires a public API change to enable state stores to report whether they need a flush or not.