Zombie TasksPipeline · medium · ~12 minKubernetes evicted three worker pods at 02:14 (node pressure). Since then, five tasks show running for 6+ hours — their processes died with the pods. Two hold slots in the critical ETL pool. Worse: when the scheduler finally reaps one and retries it, the load_ledger task DOUBLE-WRITES, because its first incarnation had already inserted rows before dying.Untangle the zombie mechanics and both layers of fix.Evidence

Evidence

Config

Checkpoints

What is a zombie task, mechanically?

Why did the retry corrupt the ledger, and whose fault is that architecturally?

Give the two-layer fix (detection + write pattern) and the interview narration.