Github user StephanEwen commented on the issue: https://github.com/apache/flink/pull/5624 The code and the tests should not rely on eventually consistent operations at all, but only on strongly consistent operations, like "read after create". That should also eliminate any need for retries. Minor comment: `System.currentTimeMillis()` should not be used for interval measurements. It is not a stable clock. All intervals should be measured in `System.nanoTime()`.