Intent Solutions · databricks-pack · editorial

Which fixes actually pay

The leak categories databricks-cost-leak-hunter surfaces, plotted by monthly dollar impact against how much confidence the evidence behind them supports. Placement is a judgement call, stated as one — not a measurement, and not a customer result.

Databricks leak categories, ranked by monthly dollars and gated by evidence confidence Quadrant chart. Six leak categories are plotted with evidence confidence on the horizontal axis, low at left, and monthly dollar impact on the vertical axis, high at top. Top right, act first, confirmed dollars: idle clusters with auto-termination disabled, which is confirmed billed spend and leads the skill's own worked example, and wasted DBUs from scheduled jobs billed on all-purpose compute, also confirmed. Top left, modeled, review before acting: over-provisioned autoscaling floors, a modeled estimate flagged at under 25 percent average CPU, and the Photon premium, detected from the SKU and held as at-risk pending a runtime-gain review. Bottom left, flagged but not ranked: untagged runaway spend, which surfaces as an attribution gap in the report footer, and oversized SQL warehouses, which the skill names in its trigger surface without a dedicated ranked query. The bottom right quadrant, confirmed but small, is empty. Placements are an editorial judgement by Jeremy Longshore of Intent Solutions, dated 13 August 2026, read off databricks-cost-leak-hunter version 2.27.0. They are ordinal positions, not measured coordinates, and no customer figures are shown. Act-first quadrant — high monthly dollars and high evidence confidence. The two confirmed categories land here. MONTHLY $ IMPACT — HIGH HERE · SETS THE RANK ORDER EVIDENCE CONFIDENCE — HIGH HERE · SETS THE GATE MODELED — REVIEW BEFORE ACTING ACT FIRST — CONFIRMED DOLLARS FLAGGED — NOT RANKED CONFIRMED, SMALL — NONE HERE Idle clusters — interactive clusters left with auto-termination disabled. Confirmed billed spend, and the top line in the skill's worked example. On a live run the ranker sorts on the customer's own figures, so any category can place first. Idle clusters AUTO-TERM = 0 · CONFIRMED Wasted DBUs — scheduled jobs billed on all-purpose compute rather than jobs compute. Confirmed, because the saving is a deterministic re-pricing delta. Wasted DBUs JOBS ON ALL-PURPOSE · CONFIRMED Over-provisioned autoscaling floors — clusters averaging under 25 percent CPU. A modeled estimate, spend times one minus CPU, not billed waste. Autoscale floor CPU < 25% · ESTIMATED Photon premium without the speedup — identified from the Photon SKU on the billed usage row. At-risk: a premium held for review against real runtime gain, not confirmed waste. Photon premium 2× SKU · AT-RISK Untagged runaway spend — spend that cannot be attributed to a team or cost centre. Reported as an incomplete-attribution flag in the report footer, not ranked as a dollar category. Untagged spend FLAGGED, NOT RANKED Oversized SQL warehouses — named in the skill's trigger surface, and the warehouse id is available on the billed usage row, but there is no dedicated ranked detection query for warehouse sizing. SQL warehouses NAMED, NOT RANKED LEGEND LEADS THE EXAMPLE LEAK CATEGORY ACT-FIRST QUADRANT

Two axes, both of them judgement. The vertical axis is monthly dollar impact — the order the skill's ranker actually sorts by, biggest number first. The horizontal axis is how much confidence the evidence carries: every line is stamped Confirmed, Estimated or At-risk, and the report's headline is split so it never sums confirmed and unconfirmed dollars under one verb. Money sets the order; confidence decides which total a figure is allowed to join.

Only the top-right pair is confirmed against the customer's own system.billing.usage: clusters left with auto-termination off, and scheduled jobs billed on all-purpose compute instead of jobs compute. The autoscaling floor is an explicitly modeled estimate and the Photon premium is a premium held for review, which is why both sit left of the confidence line whatever they are worth. The bottom-left pair are reported but never ranked — untagged spend comes out as an attribution gap in the report footer, and SQL-warehouse sizing is named in the skill's trigger surface with no dedicated ranked query behind it. The bottom-right quadrant is empty on this read: both categories the skill can confirm are ones its own cost-leak-categories.md describes as large — idle compute as one of the largest cloud-waste categories, and moving jobs off all-purpose compute as the single highest-ROI optimization in most deployments. That is a claim about typical workspaces, not a promise about yours: the ranker sorts on whatever the customer's own numbers turn out to be, so the vertical order here is the expected case, not a fixed one.

Scope check. databricks-cost-leak-hunter v2.27.0 is one of five skills in the databricks-pack. Its pipeline ranks four dollarized leak categories with a confidence tier each; the two bottom-left items on this chart are things the skill reports or names, not ranked categories, and are drawn here to make that boundary visible. Every category, confidence tier and detection query above is in the skill's own SKILL.md and references/cost-leak-categories.md. No customer result, benchmark or measured ROI appears on this page — the positions are ordinal, not numeric, and the sample dollar figures on the parent demo page are labelled there as an illustrative example, not an outcome.

Editorial placement · Jeremy Longshore, Intent Solutions · 2026-08-13 · skill v2.27.0

← Back to the cost-leak-hunter demo