Benchmarking a query optimization gave a 5x improvement on the first run, but repeated runs afterward showed almost no difference between old and new versions. Which number do you actually trust?
Accepted answer
Neither number alone tells the full story. Cold-cache performance matters for infrequent or ad-hoc queries where data won't be in memory, while warm-cache matters for hot, frequently-repeated queries. Report both and say which regime is realistic for this query's actual usage pattern.
Document the grain decision, most BI bugs turn out to be grain bugs.
Note that merge on Delta still needs unique keys defined correctly.
This matches our runbook for skewed keys.
Consider DuckDB or Polars for this size before spinning up a cluster.
How do you handle backfill without duplicating rows?
Start with the execution plan, numbers beat guesses.
Agree on the staging table swap. Atomic promote prevented partial reads.
Document the grain decision, most BI bugs turn out to be grain bugs.
Sign in to reply.
© 2026 Lakebench, operated by Hunnurji Rao. Bengaluru, Karnataka, India.
No cluster. No install. Just the tab.
