Estimate a Databricks workload's cost before you run it
The cost of a Databricks workload can be prepared before it runs: DBUs at a price per SKU, plus the cloud machines when the compute is not serverless. Here is what it is made of, how to estimate it and then check it, and what YALineo does.
What a workload's cost is made of
A Databricks workload (job, pipeline, SQL warehouse, all-purpose cluster) consumes DBUs. Databricks defines the DBU as a normalized unit of processing power. Billing is per second.
| Part of the cost | Billed by | Varies with |
|---|---|---|
| DBUs consumed | Databricks (on Azure, Microsoft) | Type and number of machines, Photon, warehouse size, hours |
| Price of a DBU | Databricks (on Azure, Microsoft) | Product (its SKU), cloud, region, plan, discounts |
| Classic compute machines | The cloud provider | Type, number, hours, payment: on demand, spot or reserved |
| Disks, storage, network | The cloud provider; storage managed by Databricks is counted in DSUs | Services chosen, cloud, region |
Each product has its own price per DBU: Jobs Compute, All-Purpose Compute, SQL Pro Compute or Serverless SQL on Azure. According to Databricks, a non-interactive workload costs significantly less on job compute than on all-purpose compute.
On serverless, the DBU price includes the virtual machines. On classic compute, the machines, their disks and the network come on top of the DBUs. Azure also bills managed disks, blob storage and public IP addresses.
The formula
DBUs per hour = DBU(driver) + workers × DBU(worker) classic compute
= DBU(size) × clusters SQL warehouse
DBU cost = DBUs per hour × hours × price of a DBU × (1 − DBU discount)
Machine cost = (price(driver) + workers × price(worker)) × hours × (1 − machine discount)
Total = DBU cost + machine cost
DBU(type), the DBUs per hour of an instance type, is published by each cloud; Photon raises it. On Azure, a Pro or Classic warehouse has the same DBUs per hour as serverless, plus the price of its machines. For a serverless job, measure the DBUs of one run on a test.
A worked example
A serverless SQL warehouse of size Small. Only its DBUs per hour come from an official page; the rest is an assumption.
| Input | Value | Origin |
|---|---|---|
| DBUs per hour, size Small | 12 | Microsoft Learn, Serverless DBU consumption by SKU |
| Hours on per month | 8 h × 22 days = 176 h | Assumption |
| Price of a DBU | $1.00 | Made-up price, round so the arithmetic is easy to follow |
| DBU discount | 10% | Assumption |
| Machines | $0 | Included in the serverless DBU price |
DBUs per month: 12 × 176 = 2,112. Cost: 2,112 × $1.00 × (1 − 0.10) = $1,900.80 per month.
Estimate before you create the workload
Hours
For a job, multiply the runs per month by the duration of one run, measured on a test. A warehouse or a cluster is paid for as long as it is on, idle time included. According to Microsoft, a SQL warehouse's auto stop comes by default after 10 minutes on serverless and 45 minutes on Pro and Classic. Databricks adds that a Pro or Classic warehouse takes minutes to start, and a serverless one a few seconds.
Size, autoscaling and Photon
With autoscaling, a cluster's workers or a warehouse's clusters vary between a minimum and a maximum. Price an average, then the minimum and the maximum. Photon raises the DBUs per hour: the cost goes down only if the run gets faster by a factor larger than the rise in the price of an hour. Databricks advises testing it on regular jobs.
Spot, reservations and discounts
- Spot: spare machines, cheaper, that the cloud can take back. Databricks recommends an on-demand driver, and spot for workloads that can afford to take longer.
- Reservations: on Azure, a 1- or 3-year commitment lowers the price of machines. The discount is lost for every hour with no machine running, and spot cannot be reserved. A part-time workload therefore also pays for the empty hours.
- DBU discounts: Databricks offers committed use contracts. On Azure, pre-purchasing units (DBCU) for 1 or 3 years saves up to 37% on DBUs, according to Microsoft. The list price does not reflect them.
The official calculators
- The Databricks Pricing Calculator asks for the plan, the cloud and the compute type, and lists the supported instance types.
- The Azure pricing calculator offers Azure Databricks. Signed in, it shows the prices of your offer.
- The AWS Pricing Calculator prices EC2 instances and their payment options, without taxes.
- The Google Cloud Pricing Calculator starts from your assumptions, and Google warns that the bill may differ.
Check the estimate after launch
Put a tag on the workload before you launch it. According to Databricks, a tag only applies to future usage, and it also reaches the cloud resources. Then compare the estimate with the measured usage:
SELECT u.usage_date,
SUM(u.usage_quantity) AS dbu,
SUM(u.usage_quantity * p.pricing.effective_list.default) AS list_price_cost
FROM system.billing.usage AS u
JOIN system.billing.list_prices AS p
ON u.cloud = p.cloud
AND u.sku_name = p.sku_name
AND u.usage_unit = p.usage_unit
AND u.usage_end_time >= p.price_start_time
AND (p.price_end_time IS NULL OR u.usage_end_time < p.price_end_time)
WHERE u.custom_tags['workload'] = 'finance-dashboards'
AND u.usage_date >= date_sub(current_date(), 30)
GROUP BY ALL
Without a tag, filter on usage_metadata.job_id (empty for a job launched on an all-purpose cluster) or usage_metadata.warehouse_id. Rows typically arrive within 12 hours. A DBU gap points to the assumption to revisit: hours, size or Photon. For classic compute, node hours are in system.compute.node_timeline, without serverless or SQL warehouses. The machine cost is in the cloud bill.
What an estimate does not tell
- The invoice: the list price leaves out your contract discount, commitments and credits.
- The real duration: a growing volume makes runs longer, and so does a spot machine taken back.
- All the cloud fees: disks, storage and network depend on the services chosen.
What YALineo does
YALineo's Calculator prices a job, a pipeline, a SQL warehouse or a cluster on Azure, AWS or GCP before you create it: DBUs, machines, total per month and per year. It works without a workspace, with the public prices of a dated snapshot bundled in the extension. Each figure says whether it is public, measured or entered (prices and their origin).
- Your prices: a discount on DBUs, one on machines, your price of a DBU (tutorial). On GCP, you enter machine prices, which Google publishes only through an API with a key.
- A new workload: hours, average workers, Photon, spot, reservation. The four ways to pay for the machines are priced side by side (tutorial).
- A measured workload: a workload of the FinOps side, read with your Databricks permissions, read only, with no YALineo server. You multiply its runs or its volume by a proportional rule (tutorial, method).
- Compare: two to four estimates side by side. Only a fully priced estimate can be named the cheapest (tutorial).
- Left out: disks, network, licenses other than Linux, taxes, credits, spending commitments, prices changed after the snapshot (full list). A workload without a price is marked "not priced", and the total is called partial.
Fictional worked cases help you learn the tool. The formulas are in the Calculator method.
Frequently asked questions
Does serverless cost more than classic compute?
Its price per DBU includes the machines; the classic one does not. Compare the total of both for the same workload and the same hours.
Where do I find my contract's price of a DBU?
system.billing.list_prices gives the list price of each SKU of your account, promotions included in effective_list. Your contract discount is not in it. On Azure, the pricing calculator shows your offer's prices once you sign in.
Do I need a workspace for YALineo's Calculator?
No. A workspace lets you start from your measured workloads and read your account's prices.
Sources
Documentation read on 6 November 2026.
- Databricks, Databricks pricing: DBU, DSU, per-second billing, prices by region and cloud, undiscounted list price, committed use contracts.
- Databricks, Pricing Calculator: the calculator's choices.
- Databricks, Best practices for cost optimization (page updated on 29 October 2026): parts of the cost, serverless, spot, autoscaling, Photon, tags.
- Microsoft, Azure Databricks pricing: price per product, DBUs per machine, Azure fees, reservations, DBCU.
- Microsoft Learn, Serverless DBU consumption by SKU (page updated on 11 September 2026): DBUs per warehouse size.
- Microsoft Learn, Create a SQL warehouse (page updated on 11 September 2026): auto stop, clusters.
- Microsoft Learn, How the Azure reservation discount is applied to virtual machines (page updated on 17 July 2026): discount lost hour by hour.
- AWS, What is AWS Pricing Calculator?: EC2, without taxes.
- Google Cloud, Google Cloud Pricing Calculator: estimates from your assumptions.
-
Databricks, Billable usage system table reference (page updated on 5 November 2026):
custom_tags,job_id, 12-hour delay. -
Databricks, Pricing system table reference (page updated on 4 February 2025):
effective_list. - Databricks, Monitor costs using system tables (page updated on 11 September 2026): the join of usage and prices.
-
Databricks, Compute system tables reference (page updated on 16 October 2026):
node_timeline.