How to Audit Your Databricks Bill from the Usage CSV (Feb 2026)

Databricks is the most opaque bill in the cloud world. The platform charges DBUs (Databricks Units) that get multiplied by a per-DBU rate that varies by cluster type, runtime, and Photon — a single number on your bill can be the result of 4 multipliers stacked.

The CSV audit untangles all of it in 60 seconds.

Step 1 — Export the usage CSV

In Databricks:

1. Admin Settings → Account → Usage
2. Pick Last 30 days
3. Download CSV

Or query the underlying system.billing.usage table directly:

SELECT workspace_id, sku_name, usage_quantity, usage_unit, list_cost
FROM system.billing.usage
WHERE usage_date >= current_date - INTERVAL 30 DAYS
ORDER BY list_cost DESC
LIMIT 100;

Step 2 — The 5 patterns

1. All-Purpose clusters running production jobs

Look at sku_name containing ALL_PURPOSE_COMPUTE vs JOBS_COMPUTE. All-Purpose DBUs cost ~2× Job DBUs. Most teams accidentally run production ETL on All-Purpose clusters because that's what they used in dev. Move to Job Compute — same exact workload, 50% less.

2. Photon on workloads that don't benefit

sku_name containing _PHOTON. Photon doubles your DBU rate but only delivers a speedup on SQL-heavy / vectorized workloads. For ML training, Python UDFs, or simple ETL — Photon is just paying 2× for nothing. We see this everywhere.

3. Idle cluster minutes

You won't find this in the basic CSV — query system.compute.clusters:

SELECT cluster_id, AVG(cpu_utilization) AS avg_cpu, SUM(uptime_seconds) / 3600 AS uptime_hr
FROM system.compute.cluster_utilization
WHERE date >= current_date - INTERVAL 7 DAYS
GROUP BY cluster_id
HAVING avg_cpu < 0.20
ORDER BY uptime_hr DESC;

Any cluster with <20% CPU average and >40 hours uptime is over-sized. Median savings: 30–60% per cluster.

4. ML Runtime on non-ML workloads

sku_name containing ML. ML Runtime costs ~15% more in DBUs. If you're using it for plain SQL/ETL, you're paying the ML premium for nothing.

5. SQL Warehouses with high auto-stop

sku_name containing SQL_PRO or SQL_SERVERLESS. Default auto-stop is 10 minutes — drop to 1 minute for non-prod and you'll cut idle DBUs 40–80%. (Yes, the cold-start is 8–12 sec. Yes, it's worth it for dev/staging.)

Step 3 — The DBU multiplier cheat sheet

Cluster	DBU multiplier
All-Purpose, Standard	1.0 (baseline)
All-Purpose, Photon	2.0
Jobs, Standard	0.5
Jobs, Photon	1.0
SQL Pro	0.55
SQL Serverless	0.7

Each multiplier is then multiplied by your list $/DBU rate (depends on plan & cloud — typically $0.40–$0.60 on AWS for All-Purpose).

So All-Purpose + Photon = 4× more expensive than Jobs Standard for the same compute. Most teams have at least one workload misclassified.

Step 4 — The 60-second pivot

In your CSV, pivot:

• Rows: sku_name
• Values: sum of list_cost

Sort descending. The top 5 SKU rows tell you exactly which cluster types are eating your bill. Then drill into each top SKU to find the offending workspace/cluster.

Step 5 — The free audit

CARTIE's free Databricks bill audit takes your usage CSV and returns:

• Your top 5 DBU leaks by cluster type
• Whether each cluster needs Photon (we cross-reference query history)
• Cluster auto-stop recommendations
• Job vs All-Purpose mis-classifications

No workspace admin access needed. CSV in, audit out.

---

Companion piece: The 6 Hidden Costs of Databricks Nobody Tells You About.