The Pushdown That Wasn'tSpark · easy · ~8 minA daily report reads one day from a year-partitioned Parquet dataset. It used to scan ~45GB; since a 'timezone fix' refactor it scans 16.4TB — the whole table. The query still returns correct results, so nobody noticed for a week. The Spark UI's scan node tells the story.Evidence

Scan Evidence

Config

Checkpoints

Why does the query now scan everything?

The fix that keeps the timezone correctness AND the pruning?

How do you make this failure loud in the future? Interview-style.