We have faced a problem with OOM errors in our dataflow job containing Kinesis sources. After investigation, it occurred that the issue was caused by too many records being consumed by Kinesis sources that pipeline couldn't handle in time.
Looking into the Kinesis connector's code, the internal queue (recordsQueue) that records are being put in the background is setup for each ShardReadersPool (created for each source being Kinesis stream). The size of the queue is set to `queueCapacityPerShard * number of shards`. The bigger number of shards, the bigger queue size. There is no ability to limit the maximum capacity of the queue (queueCapacityPerShard is also not configurable and it's set to DEFAULT_CAPACITY_PER_SHARD=10_000). Additionally, there is no differentiation on records size, so the size of data placed to the queue might increase to the point where OOM will be thrown.
It would be great to have ability to somehow limit the number of records that are being read in the background to some sensible value. At the beginning, simple solution would be to allow configuring max queue size for source at the creation of KinesisIO.