Six places where your data platform bill leaks

When the Databricks, Fabric or other data platform bill goes up, the first reaction is to change the machine or the tool. In practice, much of the cost hides in configuration details nobody looks at. These are the six places that show up most often in our assessments.

1. The machine waiting to shut down

Development environments often auto-terminate 60 minutes after the last activity. In a real case, about a third of the paid hours were just that wait. Cutting it to 15 to 20 minutes solves it without getting in anyone's way.

2. Loads running when the data does not change

In a real case, 95% to 99% of sales changes happened between 7 a.m. and 6 p.m., yet loads ran every 2 hours, 24 hours a day. Matching the frequency to when the data changes cut load time by 14%. Measure when each table changes before deciding.

3. The retry that silently doubles the cost

A step with a timeout too close to its real duration gets interrupted midway and rerun from scratch. The process ends with "success", nobody gets an alert, and the cost of that step doubles. Give the timeout some headroom and alert when the duration passes 70% of it.

4. The job that waits for itself

When a run starts taking longer than the schedule interval, each new run waits for the previous one, and the queue grows every cycle. It looks like a lack of cloud capacity, but it is the schedule. Simple rule: the typical duration must fit in the interval, with room to spare.

5. Capacity left on outside business hours

Capacity billed by the hour charges all 730 hours of the month if it never pauses. If usage happens during business hours, pausing at night and on weekends cuts most of the cost, as long as it resumes before the first refresh of the day.

6. Leftovers and forgotten charges

Disks from machines that were deleted, unused public IPs, backups of things that no longer exist. And protection lines billed under the wrong model: in a real case, changing the billing model of a storage protection service cut that line by 80%.

Where to start

Before cutting, measure: cost per layer, per load and per billing type. With the number in hand, each cut becomes a decision with a known value. That is exactly what we deliver in our data platform FinOps.