Executive Insights Walkthrough
The leadership view of your Databricks estate: what you have, how you're doing, and what to do about it
Executive Insights is the leadership view of your Databricks estate inside Lakehouse Optimizer (LHO), and the page LHO opens on. It rolls your whole tenant up to a single screen that answers three questions at a glance:
What do I have? — every workspace, region and metastore, and how your spend is split.
How am I doing? — an optimization score per lens, scored only over what LHO actually monitors.
What should I do about it? — the biggest optimization opportunities, sized and ranked, each one click from the team that can act on it.
The guiding idea is insights, not dashboarding: the page surfaces what's worth your attention and hides the rest.
How to use this guide. Each section below is a short orientation. Where a topic deserves more than a page, the section ends with a Go deeper link to a full article on just that topic — read this page top to bottom once, then return to the deep dives as you need them.
Who it's for. Two readers who share one screen: executives — a CDO, CFO, VP of Data or FinOps lead who skims the top of the page (are we healthy, and where's the upside?) — and the managers and platform owners who drill into a workspace or an opportunity and hand their team a specific, sized to-do. You don't need to be an engineer to read it.
Your tenant at a glance — the summary header
The header is the "what do I have / how am I doing" half of the page, and it is always tenant-wide — the filters lower on the page never change it. The only control that moves it is the month picker, always compared against the previous month. The current month is selectable while it is still running: it is marked Partly applied, and the month-over-month comparisons wait until the month closes — so a half-finished month is never mistaken for a drop.
Databricks Cost Plan — your commitment plans at a glance: active, exhausted or expired, with per-plan detail behind a popover. This is where the renewal conversation starts early.
Total workspaces — the estate count: workspaces, how many report telemetry, Unity Catalog coverage, regions. Workspaces are discovered from consumption data even when nobody registers them, so the total reflects the estate as the money sees it. "UC enabled" means attached to a metastore — not fully migrated.
The Optimization score — how much of what LHO monitors carries an optimization opportunity, scored per lens. Covered in its own section below.
Avg Daily Cost — a donut splitting your daily run rate into Cloud and Databricks, with the month-over-month change per slice. Links through to Cost Over Time.
The KPI band — four tiles: Cost per TB, Data Usage, Throughput, Monitored Cost. The tiles link through to Optimization Review — the evidence surface behind the numbers. The next section tells you how to read them.
The four KPIs, and how to read them
The four tiles in the KPI band are the easiest thing on the page to read — and the easiest to read wrongly. Each number is correct; the obvious inference from it is often not. The rule that covers all four:
Treat the KPI band as a smoke alarm, not a scorecard. It is built to make a large, sudden movement impossible to miss — not to tell you whether you are running a tighter ship than someone else.
Cost per TB
Your telemetry-tracked processing cost per terabyte of data moved. Read it as a trend detector at large amplitude — $18 to $25 is a conversation, a one-dollar move is not. Two cautions: it is scale-biased (fixed overhead spread over more terabytes lowers the figure on its own, so never compare it across organizations), and waste in the denominator improves it — a query that reads more data than it needs can push this number down while pushing your bill up.
Data Usage
Total data processed in the period, and the denominator of Cost per TB — read it first. It moves for three unrelated reasons: real growth, newly onboarded workspaces (a coverage artifact, not a change in the estate), or inefficient queries reading more than they need. Establish which before drawing any conclusion from Cost per TB.
Throughput
Data processed per day, computed over elapsed days so a partial month is still a fair run rate. It measures volume, not speed — a slow job and a fast job that move the same data look identical here. Nothing in this tile tells you whether anything is slow; that is the Performance lens's job.
Monitored Cost
Spend on the four asset classes LHO can write recommendations for: job compute, all-purpose compute, DLT pipelines and SQL warehouses. Read it as coverage, not performance — and never divide it by Data Usage or combine it with Cost per TB: the two cost figures are measured over different populations, and neither contains the other.
The three lenses (tabs)
Below the header, Top Optimization Opportunities organizes everything through three lenses — on screen in the order Orchestration · Performance · Cost Savings. You pick the lens; the score, the opportunity cards and the workspace ranking update to match.
Cost Savings — the cost-efficiency lens, and where most people start. Seven opportunity types, from over-provisioning and idling clusters to idle SQL warehouses.
Go deeper → Cost Savings Optimization Opportunities Guide
Performance — the speed-and-provisioning lens. Six types, from serverless candidates to skew and disk spillage. ⚠ These fixes can raise your bill — this lens optimizes latency, not cost, and says so per recommendation.
Go deeper → Performance Optimization Opportuities Guide
Orchestration — the vendor-consolidation lens, and the one that shows what other tools miss. Four types, from incomplete Unity Catalog migration to external orchestrators. It carries no savings figure, by decision — LHO sizes the exposure and stops there, rather than predicting the cost of platforms it doesn't monitor.
Go deeper → Orchestration Optimization Opportunities Guide
From a card to an owned to-do
Each lens shows a scrolling row of opportunity cards — title, affected workload count, a Cost value, a complexity pill (Low / Medium / High) and a month-over-month trend on both figures.
Read "Cost" precisely. It is the total cost of the workloads that carry that recommendation — the whole cost of those jobs, not the cost of the fix and not an estimate of what you would save. A $100k card does not mean $100k of savings; it means $100k of spend runs through workloads where LHO has found something worth looking at.
Click a card and the workspace tiles below filter to just that opportunity — "2,500 serverless candidates, but where?" answered directly. Each ranked tile carries the workspace's cost figures and workload count, and its Recommendations link opens the Recommendations page pre-filtered to that workspace and opportunity: the executive picks the theme, the platform team lands on the matching shortlist. Same data, two altitudes. The workspace ranking can be re-sorted at any time — by the cost of workloads with recommendations, by whole-workspace cost, or (on the Cost Savings lens) by potential savings.
The Optimization score
Each lens carries its own Optimization score — the share of what LHO monitors on that lens that carries no optimization opportunity. Three scores instead of one blended number, because "how optimal is my spend?", "how optimal is my performance?" and "how consolidated is my orchestration?" are different questions with different answers.
Two design principles behind it, both worth knowing because they are promises:
Only what LHO monitors is scored. If a workspace has no telemetry, its workloads are not counted — not as healthy, not as unhealthy. No opinion without data. That means the score can never be flattered by the parts of your estate LHO cannot see; the header's telemetry count tells you how much of the estate the score covers.
The score is explained, not asserted. How each figure is computed is stated on the panel itself — hover any number for its definition. No black boxes: the same explanation your team would get by asking us is built into the page.
Read the breakdown, not just the score. Under each score, the panel breaks the number down one level — by service on Cost Savings (job compute, all-purpose compute, SQL warehouses), by compute type on Performance, by opportunity class on Orchestration. The score tells you whether to look; the bars tell you where. That breakdown is the conversation.
The units differ by lens, on purpose — the panel counts what recommendations actually attach to. The Cost Savings breakdown is money: a bar per service, sized by spend. Performance counts compute instances, because that is where its recommendations live. Orchestration counts jobs and their runs — with one-time runs reported separately: a run submitted directly to the workspace by an external orchestrator is not a job you defined, and folding tens of thousands of them into a "jobs" figure would bury the number you actually manage. Wherever a figure could be read two ways, its tooltip says which it is.
Potential savings. The Cost Savings lens also carries a single estimated Potential savings figure — one number, not a range — derived from the recommendations themselves. Performance and Orchestration deliberately carry no dollar figure: performance fixes can cost money, and predicting savings on platforms LHO doesn't monitor would be a guess dressed as a number.
Asking questions with Ask Genie
Ask Genie is a natural-language way to interrogate your own LHO data: type a question in plain English — "which workspaces have the highest cost this month?", "break that opportunity down by recommendation type" — and the answer comes back in the page, with a supporting table or chart. Powered by Databricks, answering only from data you're already allowed to see.
Glossary
Optimization score — per lens, the share of monitored workloads (or spend) carrying no optimization opportunity. Scored only over services LHO has telemetry for. ⚠ Rolling out; replaces the blended Health Score.
Potential savings — a single estimated figure for the spend you could redirect by acting on cost recommendations. Produced for Cost Savings only — never for Performance or Orchestration.
Monitored Cost — spend on the four asset classes recommendations can target.
Cost per TB — telemetry-tracked processing cost per terabyte processed. Scale-biased; see the KPI section.
Data Usage — total data processed in the period; the denominator of Cost per TB.
Throughput — data processed per day. Volume, not speed.
Cost (on an opportunity card or workspace ranking) — the total cost of the workloads that carry at least one recommendation of that type. Not a savings figure.
Workload — a job, pipeline, or all-purpose-compute usage. Counts are by distinct workload, not by run, except where the page says otherwise — the Orchestration lens reports jobs and one-time runs as separate counts.
One-time run — a run submitted directly to the workspace, typically by an external orchestrator, rather than by a job you define. Counted separately from jobs wherever both appear.
Optimization opportunity — a detected, fixable issue linking to a specific recommendation.
Orchestration / vendor consolidation — the duplicated total cost of ownership from running work across parallel systems instead of consolidating onto Databricks.
Video Guides