Databricks cost cutter

We have looked at a lot of Databricks invoices. Lots of them. Most teams are not being reckless. They just turned on a serverless SQL warehouse six months ago, accepted the defaults, and never went back.

That is how a few hundred dollars a month becomes a few thousand. The defaults are not bad. They are tuned for safety, not savings.

Five settings. About thirty minutes of work if you know where to look.

1. Auto-stop at 10 minutes is generous

Serverless SQL warehouses keep themselves warm by default. Warm means ready, but also "still billing the cluster." If a developer kicks off one query, walks to a meeting, and comes back twenty minutes later, you just paid for nineteen empty minutes.

Change auto-stop to 3 minutes on dev warehouses. Five minutes on production. You can do this in the SQL Warehouse configuration panel under "Advanced options." The query wait on cold start is roughly two seconds. Nobody notices.

2. Most clusters do not need to be 4X sized

People pick warehouse size once, then forget. They pick 4X because "we already paid for it."

That is the wrong frame. A 4X warehouse costs roughly four times a 2X. If your queries run in 12 seconds on a 2X, that is fine. If they run in 6 seconds on a 4X, you just doubled your per-query cost to save six seconds that nobody was timing anyway.

Right-size them. Run the top twenty queries on Small, Medium, Large, then look at average runtime. Most teams can downgrade 40-60% of warehouses without changing user experience.

3. Forgotten warehouses are the silent killer

Open the Databricks SQL editor. Run this:

-- Warehouses idle > 15 min in the last 7 days
SELECT
  warehouse_name,
  COUNT(*) AS idle_starts_last_7d,
  ROUND(AVG(idle_seconds)/60, 1) AS avg_idle_minutes
FROM system.compute.warehouse_events
WHERE event_type = 'IDLE'
  AND event_time > current_timestamp() - INTERVAL 7 DAYS
GROUP BY warehouse_name
HAVING avg_idle_minutes > 15
ORDER BY avg_idle_minutes DESC;

Expect to find three to eight warehouses you forgot existed. Stop them. If nobody complains for thirty days, delete them.

4. Spot pools for batch jobs, on-demand for ad-hoc

Serverless is great for notebooks and ad-hoc analyst work. It is not always great for batch ETL that runs at 2am. For those, the cluster pool API with spot instances runs at roughly a third of the cost of serverless.

Worth checking if you have Lakeflow jobs doing predictable work overnight. Same compute, less money.

5. Turn on Photon if you have not

Photon is Databricks' vectorized SQL engine. For SQL workloads it gives roughly 2-3X speedup on the same cluster. Fewer seconds running means fewer DBU-hours billed.

It is on by default for serverless warehouses in current runtimes (DBR 16.4 LTS and 17.3 LTS). Verify yours is on. If you have classic warehouses still on DBR 12 or earlier, the upgrade is worth it just for Photon.

What this looks like in dollars

For a mid-size team running about twenty warehouses on serverless SQL at the starting $0.22/DBU rate:

Auto-stop 10→3 min (dev) • Monthly savings: $240-380

Right-size clusters • Monthly savings: $600-1,100

Stop forgotten warehouses • Monthly savings: $400-700

Spot pools for batch ETL • Monthly savings: $350-520

Confirm Photon is on • Monthly savings: $500-900

Realistic combined impact • Monthly savings: $2,000-3,600 /mo

Five things that are not in this list

A quick honest note on what we left out, because they show up on every listicle:

- "Set spending limits." Useful, but per-job limits create the kind of cost-tracking theater where you hit the limit once, raise it, then forget. (1/2) - "Use tags on every cluster." Yes. But tags only save money if a human looks at them. Mostly, they do not. - "Predictive IO." Real feature. Worth a separate post. - "Cluster policies." Good practice. Not a cost tip per se. - "Committed Use Discounts." The real savings, but tied to a contract and a sales call. Different conversation.

---

About Insignyx We help mid-market teams get their Databricks and Snowflake bills back under control without rebuilding the warehouse. If your monthly invoice has started to scare you, we should talk.

👉 insignyx.com