Cassandra connector FileOffsetWriter unbounded task queue causes heap exhaustion #1791
Description
FileOffsetWriter.markOffset() submits a lambda to a single-threaded executor with an unbounded LinkedBlockingQueue for every record produced to Kafka. With the default offset.flush.interval.ms=0 (AlwaysFlushPolicy), each task performs disk I/O. At high throughput the executor cannot keep up, causing millions of FutureTask objects to accumulate in the queue and eventually exhausting the heap.
Symptoms: after ~2 hours of operation, Kafka producer latency degrades from 2ms to 800ms, throughput drops, CDC logs pile up, and restart is required.