Clouds, DBUs & editions
Databricks · Databricks Fundamentals
Theory
Databricks runs as a managed service on AWS, Azure (as Azure Databricks) and Google Cloud. An account can hold several workspaces, usually one per environment or team, and each workspace is tied to one cloud region. Usage is measured in DBUs (Databricks Units): a normalized unit of processing per hour. How many DBUs a workload consumes depends on the compute type (all-purpose, jobs, SQL, serverless), its size and how long it runs.
For classic compute you also pay your cloud provider for the virtual machines, while serverless usage is billed by Databricks only. Plans (tiers) such as Premium and Enterprise decide which security and governance features you get; the old Standard tier is being retired. For example, a nightly job on job compute for one hour uses fewer DBUs than running the same code on an all-purpose cluster left on all day. Usage is visible in the system.billing.usage system table.
Why we use it
Cost questions come up in every project and in the exam. Understanding DBUs and compute types lets you explain why a job cluster or serverless job is cheaper than an idle all-purpose cluster.
When to use it
Look at DBUs when choosing compute, sizing clusters or reviewing a bill. Use system.billing.usage and the account console usage pages for real numbers. Check current pricing pages for rates; they differ by cloud, region and plan.
How it works
- Workload starts on some compute
- Databricks meters DBUs by compute type, size and time
- Is it classic compute? Yes: You also pay the cloud provider for the VMs. No: Serverless: Databricks bills all of it.
- Usage rows land in system.billing.usage
- Query or chart the usage to control cost
Syntax
DBU = Databricks Unit (normalized processing per hour)
system.billing.usage (usage_date, sku_name, usage_quantity, ...)
Account console → UsageExample
-- DBUs per product per day, last 30 days
SELECT usage_date,
sku_name,
SUM(usage_quantity) AS dbus
FROM system.billing.usage
WHERE usage_unit = 'DBU'
AND usage_date >= current_date() - INTERVAL 30 DAYS
GROUP BY usage_date, sku_name
ORDER BY usage_date, dbus DESC;Key points
- Workspaces belong to one cloud and one region; an account can have many workspaces
- Job compute and serverless jobs usually cost fewer DBUs than all-purpose compute
- Classic compute = Databricks DBUs + cloud VM cost; serverless = Databricks bill only
- Auto-termination and auto stop stop paying for idle compute
- system.billing.usage gives per-SKU usage you can query with SQL