Has anyone here looked at the new Databricks Lakehouse Replay feature?
Lakehouse Replay | Databricks on AWS
Databricks can now automaticaly take a small subset of safe, read-only workloads from your workspace and replay them against upcoming runtime versions before those versions hit production.
So if something works today but breaks on the next runtime, they can catch the regression earlier.
Honestly, this sounds pretty useful. Runtime upgrades are always one of those things that look simple on paper, but then some random query or dataframe job starts behaving differently and you're starting to scratch your head what's going on.
A few things I like:
- no setup/configuration needed
- replay runs on Databricks-managed shadow compute
- it should not impact production jobs
- customers are not billed for the replay compute
- it only compares status/metrics, not query results
I think the general idea is nice. Instead of every customer discovering regressions after upgrading, Databricks can detect some of them earlier using real workloads. That feels like something Spark platforms should maybe have had for a while.
