Hi Krishnan, Thanks for discussing this interesting scenario!
It makes me remind of a previous pending improvement of adaptive load balance for rebalance partitioner. Since the rebalance mode can emit the data to any nodes without precision consideration, then the data can be emitted based on the current backlog of partition adaptively which can reflect the load condition of consumers somehow. For your keyBy case, I guess the requirement is not only for the load balance of processing, but also for the consistency of preloaded cache. Do you think it is possible to implement somehow custom partitioner which can control the logic of keyBy distribution based on pre-defined cache distribution in nodes? Best, Zhijiang ------------------------------------------------------------------ From:Navneeth Krishnan <reachnavnee...@gmail.com> Send Time:2020年9月23日(星期三) 02:21 To:user <user@flink.apache.org> Subject:Adaptive load balancing Hi All, We are currently using flink in production and use keyBy for performing a CPU intensive computation. There is a cache lookup for a set of keys and since keyBy cannot guarantee the data is sent to a single node we are basically replicating the cache on all nodes. This is causing more memory problems for us and we would like to explore some options to mitigate the current limitations. Is there a way to group a set of keys and send to a set of nodes so that we don't have to replicate the cache data on all nodes? Has someone tried implementing hashing with adaptive load balancing so that if a node is busy processing then the data can be routed effectively to other nodes which are free. Any suggestions are greatly appreciated. Thanks
