Home › Learn › Databricks compute and cost › Dividing the bill

Databricks compute and cost · data engineering

Dividing the Databricks bill: the chargeback architecture

Force tags, record usage, roll it up: the architecture for charging Databricks cost to business units and environments, the four workspace layouts compared, and the per-query gap.

Cluster anatomyWarehouses and serverlessChoosing computeCompute settingsCost by workloadLogs and visibilityPhoton and cachesHow billing worksDividing the billSystem tablesSharing and regionsBackup architectureCost-saving hacks

Prices, defaults, limits and preview status change often, and dollar figures here are illustrations. Check the current Databricks documentation and price list before quoting a number.

Dividing the bill

Best architecture for dividing the bill

Chargeback works when three things are true: every unit of compute is owned by exactly one BU and environment, that ownership is written down as a tag by a rule (not by hand), and one rollup table turns tags into money.

1. FORCE the tags Cluster policy: tags requiredServerless budget policy per BU, per environmentWarehouse tagsJob tags, run as service principal per BU Users cannot skip a tag. 2. RECORD system.billing.usageDBUs, tags, workspace,cluster, warehouse, job system.billing.list_pricesprice per DBU per SKU Cloud cost exportVMs and storage, classic only 3. ROLL UP and SHOW One small daily jobusage × price, join cloud cost,write to a chargeback table Dashboard by BU / env / product Budgets and alerts per BU Untagged spend gets its ownline so it is visible. Shared costs (platform workspace, pools, admin warehouses, storage)Put them in their own tagged workspace and split them across BUs by a stated rule, for example proportional to each BU's own spend. Never hide them in one BU.
Prevention first (policies make tags mandatory), recording second (system tables), allocation third (a rollup you own). Fixing untagged spend after the fact is much harder.
LayoutHow the bill splitsStrengthWeaknessVerdict
One workspace, tags onlyBy tag on clusters, warehouses, budget policiesSimplest to set upOne missed tag poisons the bill. Weak access isolation between BUs.Drop for more than one BU
Workspace per environment (prod, dev), BU by tagWorkspace gives environment, tag gives BUProd is isolated from devBUs still share a workspace, so BU boundaries rest on tagsPursue for small organisations
Workspace per BU and environment (BU1-Prod, BU1-Dev, BU2-Prod, BU2-Dev) + catalog per BU, tags as a second checkWorkspace id is in every usage record, which makes the split almost automaticCleanest bill, cleanest access, simple budgets per workspaceMore workspaces to manage and more idle warehouses. Needs automation (Terraform).Pursue as the default for several BUs
Separate account (and cloud account) per BUCompletely separate invoicesHard financial and legal separationNo shared catalog or governance, duplicate admin, harder to see group-wide spendDrop unless regulation or a subsidiary demands it

Rules that make any layout work: a single naming standard for tags (bu, env, cost_center, project, owner); production jobs run as service principals named per BU, never as people; one SQL warehouse per BU, environment and workload class so usage is not mixed; set budget policies and alerts, not just reports; and review the untagged line every week.

The per-user and per-query gap. Billing is recorded per warehouse or cluster, not per query or per person. If five teams share a warehouse, you can only split its cost by estimating, using each team's share of query time from query history. That estimate is fair but is not an invoice figure. This is the strongest argument for separate warehouses per BU.

FAQ

Questions people ask

What is the best layout for several business units?

A workspace per business unit and environment, a catalog per business unit, and mandatory tags as a second check. It gives the cleanest bill and the cleanest access, at the cost of more workspaces to automate.

Can I get cost per query or per user?

Not from billing. Billing stops at the warehouse or cluster. You can estimate shares from query history, which is why separate warehouses per BU are best.

Next

Where to go next

Related on this site: Tagging best practices · Cost and FinOps for the platform

All Databricks compute and cost pages · All data engineering · Databricks on this site