Executor OOM on a Skewed JoinSpark · medium · ~12 minThe sessions ⋈ devices join dies with executor OOMs — but only ever on one or two tasks; the rest of the stage breezes through. The team's first instinct was doubling executor memory. It failed the same way, on the same task, a week later.Decide between the two real fixes — the evidence points at the cheaper one.Evidence

Stage Evidence

Profile Detail

Config

Checkpoints

Why did doubling executor memory fail?

Given the evidence, which fix is right — and why?

When WOULD salting be the right answer instead? And what would you do about 'unknown' device_ids longer-term? Interview-style.