Legacy (SDF wrapper-based) KafkaIO reader degrades when Kafka poll latency is hign

Beam has two KafkaIO source implementations now:

• a modern one implemented as a Splittable DoFn (SDF), and
• a (deprecated) legacy one implemented as an SDF wrapper over an UnboundedSource and KafkaUnboundedReader classes.

We found that the legacy KafkaIO source can not provide good throughput when the latency of calls to Kafka Consumer.poll becomes high. The degradation is very sharp: a pipeline that drops elements immediately after reading them from source was only able to read about 100-1000 qps per Kafka partition. The Kafka cluster was overprovisioned but was in a remote network and had poll latency about 30ms.

First problem that may be addressed in the scope of this bug is that there's very little visibility into the Kafka source now. We had to add extra logging to understand the issue with the pipeline above, or even see the poll latency.

We believe that the cause of throughput degradation is poor choice of the RECORDS_DEQUEUE_POLL_TIMEOUT and RECORDS_ENQUEUE_POLL_TIMEOUT especially the former one which is now 10ms.

These are timeouts for popping and pushing elements from/to the availableRecordsQueue .  This is a synchronous queue (i.e. blocking, without buffering) used to hand records fetched from Kafka between two loops:

• The consumerPollLoop that polls data via a Kafka Consumer if there's no pending data already, and offers it to the availableRecordsQueue otherwise. It also does offset checkpointing but this is irrelevant to our case.
• The beam UnboundedSourceAsSDFWrapperFn message processing loop. It's a bit complicated, but the important part is that it would call the nextBatch function repeatedly until an attempt to fetch an element from the avaliableRecordsQueue times out. After the timeout, it returns the control to the worker and it may take relatively long time until the loop is scheduled again.

This is what we think is happening when the poll latency is high:

• consumerPollLoop fetches data bundle from Kafka via poll() and offers it to the avaliableRecordsQueue
• message processing loop fetches bundle from avaliableRecordsQueue and unblocks the consumerPollLoop
• consumerPollLoop calls poll() again
• message processing loop completes processing the bundle BEFORE the poll() call above completes, and tries to fetch next bundle from avaliableRecordsQueue.
• fetch from avaliableRecordsQueue has a very short timeout (10ms) and if it expires before the pending poll() in the consumerPollLoop completes the message processing loop will believe there's no fresh data in Kafka and exit. All the time until the message processing loop is rescheduled is wasted.