What Is a Databricks DBU?
Post · List rates illustrative · Updated with site JSON 2026-09-24
A Databricks Unit (DBU) is Databricks’ normalized measure of platform compute consumption — not a cloud VM, and not a fixed dollar by itself. Dollars appear when you multiply DBUs by the $/DBU list (or contract) rate for your SKU; metering for classic compute is typically accrued continuously (commonly described as per-second granularity billed to the hour/policies your account uses), so short runs still count.
If you only remember one sentence for finance: SKU drives $/DBU; runtime and cluster shape drive how many DBUs you burn; the cloud still bills the machines underneath.
DBU in plain language
Think of a DBU as a “platform usage token.” Different workload types (Jobs Compute, All-Purpose / interactive, SQL Warehouse, and various serverless offerings) consume DBUs at different rates and publish different list prices per DBU. A bigger instance type generally consumes more DBUs per hour than a smaller one on the same SKU family — the product maps hardware capability into DBU/hour.
That mapping is why two clusters with the same wall-clock hours can produce very different DBU totals. It is also why “we ran for one hour” is not enough information for a cost estimate.
FAQ: Is metering per second?
Yes, in practical terms classic Databricks compute usage is metered with fine granularity (often described as per-second), then rolled into your bill according to product rules. You do not need to “fill” a full hour to owe DBUs for a five-minute job — you owe for the seconds (or equivalent) the compute was allocated, subject to how that product meters startups, minimums, and serverless packing.
Education implication: short, frequent jobs are not “free because they are under an hour,” but they also should not be costed as if every run were a full reserved hour of All-Purpose. Use actual duration × nodes × DBUs-per-node-hour for estimates, then label the result as an estimate.
FAQ: What drives $/DBU?
Your SKU (and cloud + region + discounts) drives $/DBU. On this site’s illustrative list file (/data/pricing/rates.json, last updated 2026-09-24), Jobs is shown near $0.15 / DBU while All-Purpose is near $0.55 / DBU across the sample clouds. Those figures are list education rates, not a quote and not Photon- or serverless-specific.
Contract customers often pay less than list. When you know your rate, override $/DBU in the calculator instead of trusting the JSON.
Worked micro-example
Suppose a Jobs cluster with 4 workers, each mapped at ~1.5 DBUs per node-hour (education default — replace with your instance’s real DBU rating). The job runs 20 minutes (1/3 hour).
- Node-hours ≈ 4 × 1/3 ≈ 1.33
- DBUs ≈ 1.33 × 1.5 ≈ 2.0
- DBU $ at $0.15 list ≈ ~$0.30 for that run
That ~$0.30 is only the Databricks platform line. The four VMs (plus driver) still incur cloud list/spot/RI charges for those ~20 minutes — see DBU + VM total cost. Monthly, multiply by run count.
Using the sample retail_orders table keeps the job tiny so you can focus on the arithmetic; production facts change hours and nodes, not the formula.
DBU vs “instance hour” vs “cluster hour”
- Cluster hour — wall clock while the cluster exists
- Node hour — cluster hours × number of nodes (workers, and often driver)
- DBU — node hours × DBU rating for the instance/SKU configuration
- DBU $ — DBUs × $/DBU for that SKU
People mix these terms in Slack and under-budget. Write the unit on every slide.
What a DBU is not
- Not a guarantee of performance
- Not inclusive of cloud VM / disk / network charges on classic compute
- Not the same list price across Jobs vs All-Purpose vs SQL
- Not automatically equal to your discounted contract rate
Serverless products package infrastructure differently; read current product docs before applying classic VM math. One sentence caveat: serverless still has a cost model — it is just not “DBU + you pick the EC2 type” in the same way.
How to estimate without overclaiming
Always label outputs: list price, estimate, rates as of DATE. Do not promise “X% cheaper than Y.” Point stakeholders at official pricing pages and the bill. Our JSON cites public list sources for education; they can drift.
For a full Jobs walkthrough with the retail sample, read the TCO guide.
Takeaway
DBUs measure Databricks platform consumption; SKU sets $/DBU; fine-grained metering means short jobs still accrue usage. Add infrastructure cost separately for classic clusters, and verify rates before budgeting.