Hi, As for most pipelines, our flink pipeline start with parsing source kafka events into POJOs. We perform this step within a KafkaDeserizationSchema so that we properly extract the event itme timestamp for the downstream Timestamp-Assigner.
Now it turned out that parsing is currently the most CPU intensive task in our pipeline and thus CPU bounds the number of elements we can ingest per second. Further splitting up the partitions will be hard as we need to maintain the exact order of events per partition and would also required quite some architectural changes for producers and the flink job. Now I had the idea to put the parsing task into ordered Async-IO. But AsyncIO can only be plugged in into an existing Stream, not into the deserialization schema, as far as I see. So the best idea I currently have is to keep parsing in the DeserializationSchema as minimal as possible to extract the Event timestamp and do the full parsing downstream in Async IO. This however, seems to be a bit tedious, especially as we have to deal with multiple input formats and would need to develop two parsers for the heavy load once: a timestamp only and a full parser. Do you know if it is somehow possible to parallelize / async IO the parsing within the KafkaDeserializationSchema? I don't have state access in there and I don't have a "collector" object in there so that one element as input needs to produce exactly one output element. Another question: My parsing produces Java POJO objects via "new", which are sent downstream (reusePOJO setting set) and finally will be garbage collected once they reached the sink. Is there some mechanism in Flink so that I could reuse "old" sinked POJOs in the source? All tasks are chained so that theoretically, that could be possible? Best regards Theo
