Performance Forensics — Week 2 CheckSymptom to mechanism: joins, memory at both ends, codegen, and shuffle physics. Every question is answerable from the Week 2 brief.

A join fails with executor OOMs and task retries. The driver is healthy, and the failed tasks show no sort phase. Which join algorithm is implicated?

A stage slowed ~3× after a column was reimplemented with a Scala UDF. What Spark UI evidence would CONFIRM the codegen-fallback diagnosis?

A fleet shows ~50% aggregate memory utilization AND hundreds of executor OOMs per day. What single root cause explains both numbers at once?

An automated memory-tuning loop applies exactly ONE sizing policy per job run, never stacking changes. What does this rule buy?

Magnet's push-merge shuffle deliberately refuses to merge blocks above a size threshold. Why?

In a partition-owned remote shuffle service, every record carries its map task's attempt id. What failure would corrupt results if this were missing?