Use Case

Databricks Cost Optimization

Q: Doesn't Databricks already optimize costs?

Native tools give you visibility and storage upkeep. Cost dashboards and budgets show spend; Predictive Optimization maintains tables. None of them change a cluster or warehouse config. Altimate acts on compute, the part that dominates the bill, and every action carries a snapshot and an automatic rollback.

Q: What access does Altimate need?

A scoped service principal with read-only metadata access for analysis. Write permission exists only on the specific jobs, warehouses, and clusters you enable Auto Tune for, and it changes configuration only: never code, schedules, or data. Unity Catalog is required.

Q: How are savings measured?

Per run, from your billing data, at your contracted rate rather than list price. Realized savings appear next to each tuned job, warehouse, and cluster, and roll up to the Summary page ledger.

Q: What about serverless?

Serverless SQL warehouses benefit most: they can stop within minutes of going idle and resume in seconds. Serverless job tasks have no cluster to configure, so Auto Tune skips them while cost tracking still covers them.

Q: What happens if a change slows something down?

Auto Tune watches every touched job, warehouse, and cluster continuously. A failed run, or runtime beyond 1.5 times the recent baseline, triggers Backoff: the prior configuration is restored automatically and the event lands in the history.

Auto Tune agents right-size your job clusters, SQL warehouses, and all-purpose compute on their own — approval-gated, checked against your SLAs, and rolled back automatically if anything slows down. For everything else, a co-pilot does the heavy lifting: sizing each opportunity and routing it to the right owner to approve.

Book a Demo

20–35%

typical cost reduction

100+

saving opportunities surfaced

<30

days to proven value

01 · Auto Tune

Autonomous Compute Savings, SLA Safe

Auto Tune reads how your job clusters, SQL warehouses, and all-purpose clusters actually run, then moves each to a cheaper setup on its own. Approve it once; it applies between runs, monitors continuously, and rolls back the moment anything slows down.

Jobs, warehouses, and all-purpose compute

Saves both DBU and cloud costs

SLA-safe, approval-gated, auto-reverts

Every automated change runs the same loop. No exceptions, no surprises.

01

Approval-gated — you choose the jobs, warehouses, and clusters it may touch.

02

Policy-checked — changes clear your SLA rules before they apply.

03

Atomic — one snapshot, one change, applied only between runs.

04

Watched — continuous monitoring against the prior baseline.

05

Reversible — failure or degradation triggers automatic rollback.

Watch a real pass: analysis, gate checks, the applied change, the watch window that follows, and the audit trail every event lands in.

02 · Spark & job intelligence

Spark Waste, Found and Fixed

Slow shuffles, oversized executors, steps that run one after another when they could run together. Altimate reads each run, prices it, and recommends the fix — so a job creeping up on cost is caught weeks before month-end, with a concrete change already in hand.

Finds hidden Spark waste

Alerts when cost or runtime jumps

Recommends the fix, not just the flag

A Spark run flagged with a 1.0 GB spill and a 6.4x skew ratio, above three recommendations: eliminate the disk spill in the sort-merge join, reduce data skew on the join key, and cut GC pressure in the hash aggregate — each with its cost and time savings — Altimate reads each Spark run — here flagging a 1.0 GB spill and a 6.4× skew — then hands back concrete fixes: eliminate the disk spill, rebalance the skewed join key, ease GC pressure. Each with the cost and time it saves.

03 · AI/ML spend intelligence

The Fastest-Growing Line on the Bill Is the Least Watched.

Model serving, vector search, AI gateway, and Genie — the fastest-growing spend on the platform. Altimate attributes every dollar by service, team, and day, and flags endpoints serving nobody.

AI/ML Services cost dashboard showing Model Serving, Vector Search, and AI Gateway spend over 28 days — Every dollar of AI/ML spend attributed by service and by day across all workspaces. Teams finally know where the budget is going, instead of stitching it together from billing exports after the fact.

3 idle model-serving endpoints. 30 days. Zero requests. $19,800/month in invisible waste.

We flagged it. They fixed it. $240K saved annually. Zero disruption.

Opportunity card recommending decommissioning three idle model serving endpoints for roughly twenty thousand dollars per month — $20K/month. Low effort. First detected 5 days ago. Altimate doesn't wait for a quarterly review to find invisible spend. It surfaces the opportunity, sizes the saving, and hands it to the right engineer to action.

04 · Discover

Every Opportunity Sized, Owned, and Tracked to Done.

Spot-instance misconfigurations, wrong instance families, notebooks parked on all-purpose clusters, resources idling for days. A co-pilot does the heavy lifting, turning each into an opportunity with a dollar value, an effort score, and a name next to it.

Links straight to the resource

Filter, group, and export

Tracked until it's done

Discover page listing Databricks cost opportunities with money savings, time savings, effort, status, and owner columns — The Discover queue: each row carries money savings, time savings, effort, status, and an owner.

From finding to fix: how an opportunity gets explained, assigned, and closed.

05 · Chargeback & forecast

Know Which Team Spent What. See Next Quarter Early.

Spend split by team first — then by workspace, SKU, product, and user — and projected forward. Every dollar rolls up to an owning team, so budget conversations happen before the invoice arrives, not after.

Spend by team and user

Broken down by SKU and product

See this quarter and next

Breakdown dashboard: total cost of $603K across six workspaces, with cost broken down by product — SQL, Jobs, all-purpose, model serving, DLT, and vector search — each with DBUs and percent of total — Total spend split by workspace and by product — SQL, Jobs, all-purpose, model serving, and the rest — each ranked by share of the bill.

Users dashboard ranking the top Databricks users and service principals by query cost, with DBUs, query counts, success rate, and average duration per user — Top users and service principals ranked by query cost, with DBUs, query volume, and success rate. Chargeback stops being a spreadsheet exercise.

Team sport

Cost Optimization Is a Team Sport.

Autonomous savings and the opportunities your team closes land on one Summary page — measured in dollars and engineer-hours, each with an owner. Cost optimization stops being only the platform team's job.

AUTONOMOUS

Auto Tune agents apply the change themselves. Gated, monitored, reversible.

ASSISTED

A co-pilot does the heavy lifting — sizing each opportunity in dollars, scoring the effort, and naming an owner. Your team gives the final go-ahead.