Page History

...

1. All but the final result for windows. You can use suppress get exactly one final result per window/key for windowed computations. This includes both time and session windows.
   1. This feature requires adding a "grace period" parameter for windows.
2. Intermediate Events. You can use suppress to wait for more updates on the same key before emitting them downstream. There are two main configuration knobs to tune here:
   1. How long to wait for more updates before emitting. This is an amount of time, measured either from the event time (for regular KTables) or from the window end start (for windowed KTables), to buffer up each key before emitting them downstream.
   2. How much memory to use for buffering events (and what to do when it fills up). Suppress needs to remember the last observed event for each key, which takes memory. You can configure the amount of memory either by number of keys or by raw heap size. You can configure what to do when the buffer runs out of memory:
      1. Emit when full. This option is basically "best effort" suppression. If suppress runs out of memory, it will emit the oldest event.
      2. Spill to disk. This option is for when you need suppression to be strict. If the operator runs out of memory, it will allocate a RocksDB database and begin storing new records there. This may be complicated to implement, and consequently might be a future addition.
      3. Shut down. There is an option to run with the assumption that your "live set" of suppression events will fit within memory. If this assumption turns out to be false, the application will run out of memory and shut down. This may sound strange, but it's no different than using an InMemoryKeyValueStore or even an ArrayList. In the case of suppress, you can either leave the memory buffer unconstrained, which will cause an application crash if it runs out of memory, or you can configure it with a memory constraint and tell it to shut down gracefully when it runs out. Either one is safe from a correctness standpoint, but the latter will result in faster recovery. 

...