The Spark Bill — Week 3 CheckTask ledgers, spot economics, and acceleration boundaries. Every question is answerable from the Week 3 brief.

A job scans 6 TB of well-sized Parquet with spark.sql.files.maxPartitionBytes at 512 MB. Roughly how many input tasks does the planner create?

A write uses partitionBy on a column with 30 distinct values, and the writing stage runs 400 tasks with rows for every value spread across all tasks. What is the worst-case output file count?

On a spot-heavy Spark cluster, which single node loss is fatal to the whole application rather than recoverable?

Beyond surviving spot interruptions, what does moving shuffle to shared storage (disaggregated shuffle) unlock for cluster economics?

Which job profile is the strongest candidate for GPU or native vectorized acceleration?

A native-execution plugin reports a 2× average speedup but an 8× best case across queries. What mechanism explains the wide variance?