DBU by Cluster (7d), Databricks

​

At a glance

Databricks Units (DBU) consumed over the last 7 days, attributed per cluster and ranked highest first. DBU is the normalised unit Databricks bills on: every cluster, SQL warehouse, and job-compute resource burns DBU per hour at a rate set by instance type and workload (Jobs, All-Purpose, SQL). This card answers the question a platform owner asks every Monday: “Which clusters spent our money last week, and is the ranking what I expected?” A single cluster quietly dominating the bar chart is the most common root cause of an unexplained invoice.

​

Calculation

DBU(cluster, 7d) = Σ usage_quantity
                   WHERE usage_date >= today - 7
                   AND usage_metadata.cluster_id = <cluster_id>

1. All SKUs are included. A single cluster can burn DBU under more than one sku_name (for example ENTERPRISE_ALL_PURPOSE_COMPUTE for interactive use and ENTERPRISE_JOBS_COMPUTE if it also runs scheduled jobs). The card sums across SKUs per cluster so the bar reflects total spend on that cluster, not one workload type.
2. DBU, not currency, is the native unit. The card displays DBU because the $/DBU rate varies by SKU and by contract. If your connector profile carries a blended rate, an optional currency overlay is shown; otherwise DBU is authoritative and directly reconcilable against the Databricks account console.
3. Photon and serverless are counted. Photon-accelerated clusters burn DBU at a higher multiplier; serverless SQL and serverless jobs report their own SKUs. All appear in the same ranking so a “cheap looking” serverless workload that is actually expensive is not hidden.

​

Worked example

1. The nightly ETL is the biggest line, and that is expected. prod-etl-nightly topping the chart is healthy: it is the workload that produces the business’s core tables. The owner’s job here is not to cut it but to confirm it has not crept. Comparing against last week’s snapshot, it was 3,950 DBU, so a 7% rise. Worth a glance but not an alarm; pair with Avg DBU per Job Run to see whether the rise is more runs or heavier runs.
2. The shared all-purpose cluster is the suspicious line. analyst-shared-ap at 3,880 DBU is almost as expensive as the entire production ETL, and it is interactive, meaning analysts leave it running. Cross-reference Idle Cluster DBU Wasted (24h): if a large slice of that 3,880 was burned with no active query, the fix is a tighter auto-termination setting (drop from 120 minutes to 30) rather than a smaller cluster.
3. The “Unattributed” bar is a governance gap, not a workload. 520 DBU with no tag means some compute is being created without the team’s tagging policy applied (often ad-hoc clusters spun up from notebooks). It is only 4.6% of spend today, but untagged compute is the seed of next quarter’s unexplained invoice. The action is to enforce a cluster policy that requires the cost_centre / job_id tag.

Quick FinOps framing for the week:
  Production (ETL + ML, ranks 1/3/4):  6,840 DBU  → core, justified
  Interactive (analyst shared, rank 2): 3,880 DBU  → review idle %
  Untagged (rank 5):                      520 DBU  → close the tagging gap
  -----------------------------------------------------------------
  Optimisation headroom this week:    ~1,500-2,000 DBU if idle + untagged tightened

​

Sibling cards to read alongside

​

Reconciling against the source

Account console → Usage is the authoritative billing view. Filter by the 7-day range and group by cluster to match this card. System tables: query system.billing.usage directly in a SQL warehouse, grouping by usage_metadata.cluster_id and summing usage_quantity for the same dates. This is the exact source the card reads. Cluster details: the Compute page shows each cluster’s DBU/hour rate and current state to sanity-check the per-bar attribution.

SELECT usage_metadata.cluster_id,
       SUM(usage_quantity) AS dbu_7d
FROM   system.billing.usage
WHERE  usage_date >= current_date() - INTERVAL 7 DAYS
GROUP  BY usage_metadata.cluster_id
ORDER  BY dbu_7d DESC;

​

Known limitations / FAQs

​

Tracked live in Vortex IQ Nerve Centre