Progress
1 visited · 0 revised
Chapter
10 / 10
On this page
Q/A
Q/A
How do you handle a Flink job falling behind when consumer lag keeps growing?
What I would say
I treat this as a throughput-plus-safety-window problem, not just a restart problem. First I alert on lag growth, processing-time delay, checkpoint duration, and watermark stall so I know whether the issue is backpressure, skew, bad state growth, or an upstream traffic spike. Then I scale parallelism using historical throughput and partition math, rebalance hot keys if one partition is dominating, and protect Kafka retention as the short replay buffer while the job catches up. If the lag threatens to run past that retention window, I stop pretending Kafka alone is enough and switch to an Iceberg-backed replay or correction path so correctness is preserved instead of silently losing late data.
If they push further
What would you inspect first to separate a hot key from generic under-provisioning?
What not to say
I would just restart the Flink job and hope it clears the lag.
Last reviewed June 2026 · By Prasoon Parashar
Numbers are interview assumptions, not real Netflix internal figures.