Databricks FinOps Cost Accelerator
Multi-workspace Databricks cost visibility from system tables: SKU/job/user attribution, monthly compute-type trends, and optimization signals. Works on any workspace.
Product Preview
Try the interactive reader and demo tools below, or get the full product with all content unlocked.
⚙ Try Demo ToolsFile Structure 18 files
Documentation Preview README excerpt
Databricks FinOps Cost Accelerator
Multi-workspace Databricks cost visibility built entirely on system tables —
SKU/job/user attribution, monthly compute-type trends, and reserved-instance
utilization. Works on any Databricks workspace with zero organization-specific
dependencies: everything is driven by system.billing.usage,
system.billing.list_prices, and system.lakeflow.jobs.
Why it exists
Databricks bills at the DBU/SKU level, but the questions leadership asks are
"which team, which job, which workspace, and where can we cut 20%?" This
accelerator answers those in a single databricks bundle deploy — no custom
cost pipeline to build.
What's inside
- A Databricks Asset Bundle — one daily job, five parameterized SQL tasks.
- Aggregate tables: daily usage by SKU/workspace, daily job costs (with job
names resolved), monthly compute-type breakdown, per-principal attribution,
and reserved-instance utilization.
- A dependency-free analysis library (
lib/cost_analysis.py) that mirrors
the SQL rollups in pure Python — SKU classification, cost rollups, top jobs,
RI utilization, and an all-purpose-compute right-sizing signal.
- A unit-test suite (
tests/, 9 passing tests) that runs offline.
Quickstart (30 minutes)
# Verify the analysis logic offline first (no Databricks needed):
pip install pytest
pytest tests/ -v
# Then deploy:
databricks bundle validate -t dev -p <your-profile>
databricks bundle deploy -t dev -p <your-profile>
Set catalog_name, schema_name, and warehouse_id in databricks.yml.
Ensure the deploying principal can read system.billing.*.
The right-sizing signal
The classic Databricks overspend pattern is interactive all-purpose compute
running scheduled work that belongs on cheaper Jobs/Serverless compute.
rightsizing_signal() (and the RI utilization table) surface exactly that, with
an estimated saving. Typical finding: 10–30% reducible spend.
Contents
databricks.yml resources/jobs.yml
src/01_daily_usage_by_sku_workspace.sql
src/02_daily_job_costs_by_workspace.sql
src/03_monthly_aggregates.sql
src/04_monthly_usage_by_user.sql
src/05_reserved_instance_utilization.sql
lib/cost_analysis.py tests/test_cost_analysis.py
guide/ guides/finops-methodology.md
... continues with setup instructions, usage examples, and more.