Databricks is powerful — and easy to overspend on. These moves cut DBU burn fast while keeping performance.
Quick Wins
- Job clusters > all-purpose: per-job spin-up; kill after run.
- Autoscaling min=0: allow full scale-to-zero on low-traffic jobs.
- Spot/preemptible mix: 50–70% workers spot where workloads tolerate retries.
- Photon runtimes: enable for SQL/ETL boosts at same DBUs.
- Delta caching + Z-ordering: reduce IO and scan costs on hot tables.
Guardrails
- Max cluster size caps per workspace; tag owner/cost-centre on clusters & jobs.
- Auto-terminate after 10–15 min idle; block “no-limit” autoscaling.
- SQL endpoints with size tiers; archive/compact small files weekly.
Watch These Metrics
- DBU per job/run: trend by pipeline
- Cluster idle %: per workspace
- Failed-due-to-preemption: keep < 3% per week
- Small-file ratio: files < 128MB / total
Need a Databricks tune-up?
We’ll baseline, apply quick wins, and hand over guardrails in two weeks.
Optimise my workspace