Benchmarks you can check yourself, and a guide to running Goldilox Insights on your own tables.
If you want to check our numbers before you trust them, this is the run they came from, uncut. The tables start with no clustering, the 99 TPC-DS queries go through once, Goldilox Insights reads the workload and picks keys, then the same queries run again. Data scanned fell 81%, runtime 51%, credits 50%.
Watch walkthrough → GuideWhat installing Goldilox Insights actually involves. You get it from the Snowflake Marketplace, point it at the tables you care about, and review the keys it suggests before anything is applied. Most teams are done in about 15 minutes, and no data leaves their account.
Read guide → BenchmarkIf you've moved tables to Iceberg, clustering still matters there. An unclustered Iceberg table reads just as much as an unclustered native one. We ran TPC-H SF100 on a Snowflake-managed Iceberg table and let Goldilox Insights pick the keys. Pruning reached 32% and data scanned fell 40%, close enough to native to call it even.
Read article → BenchmarkOur first published benchmark, from April 2025, kept here for the record. On 10 TB of TPC-DS data the keys Goldilox Insights chose lifted partition pruning from 41% to 59% and cut query time by 12%. The engine has moved on a long way since. The walkthrough above shows where it stands now.
Read article →