docs/PRD.md

PRD: databricks-cost-leak-hunter

Author: Jeremy Longshore (Intent Solutions) Date: 2026-07-07 Status: Active

Backfilled 2026-07-07 from the shipped skill (SKILL.md v0.1.0, eval-spec.yaml, and the pack's design-review record) to the templates/skill-docs/ submission standard — the first first-party skill held to the pack/flagship tier of the doc matrix. Refs #984.

Problem

Databricks bills climb without an explanation a finance owner can act on. The waste hides in four recurring configurations: clusters that never auto-terminate and bill around the clock for idle compute; scheduled jobs running on All-Purpose Compute ($0.55/DBU) instead of Jobs Compute ($0.15/DBU) — 2–3× the cost for identical work; clusters sized for peak that idle below 25% CPU; and the ~2× Photon DBU premium paid on jobs it doesn't accelerate. The native billing data can prove all four, but it lives in system.* tables behind a metastore-admin grant chain and speaks in DBUs — so cost reviews either stall on access, or arrive as engineering jargon a CFO can't act on, or fall back to generic industry waste estimates that a skeptical reader can dismiss.

Target users

User	Context	Primary need
CFO / FinOps owner	Monthly spend review, or the Databricks bill just spiked	A dollar-ranked report readable in ~90 seconds and actionable without an engineer to translate
Data-platform engineer	Asked "why is our Databricks bill so high?" and needs defensible numbers, not vibes	Deterministic detection SQL over system.billing.usage plus a ranked list of one-config-change fixes
Databricks workspace admin	Recurring cost-hygiene sweeps (idle clusters, wrong-SKU jobs, oversized floors, Photon usage)	A repeatable detect → compute → rank → report pipeline with confirmed-vs-estimated labeling

Success criteria

Criteria 1–3 are enforced verbatim by the skill's eval-spec.yaml judge criteria.

1. Triggers on Databricks cost questions ("why is my databricks bill", "find wasted spend", "cost leak") and does not trigger on unrelated prompts — eval criterion triggers-on-cost-question (blocker) plus the two should-not-trigger control cases.
2. A CFO with no Databricks engineering knowledge can read the output in ~90 seconds and say "we are wasting $X here, fix it" without an engineer — eval criterion produces-cfo-grokkable-report (blocker).
3. Confirmed and estimated/at-risk dollars are never summed under one figure, and every leak carries a Confidence label (Confirmed / Estimated / At-risk) — eval criterion splits-confirmed-vs-estimated (regression-critical), with confirmed figures attributed to system.billing.usage (dollars-from-billing-not-estimates).
4. A workspace missing the billing grant chain is reported upfront with the exact missing GRANT statements — Step 1 fails fast instead of dying mid-analysis (eval criterion checks-grant-chain-upfront).

Functional requirements

• FR-1: Verify the metastore-admin grant chain on system.billing before any analysis; on failure, report the exact missing grants verbatim and stop.
• FR-2: Detect the four named leak categories with SQL over system.billing.usage joined to system.billing.list_prices via the CLI Statement Execution API: idle clusters (auto_termination_minutes = 0), scheduled jobs on ALL_PURPOSE, overprovisioned clusters (<25% CPU from system.compute.node_timeline), and the Photon premium (sku_name ILIKE '%PHOTON%').
• FR-3: Corroborate each leak's live configuration through databricks-workspace-mcp (clusters_get, clusters_events, clusters_list, instance_pools_list, pipelines_get); if the MCP server is absent, still produce the dollar figures and accept pasted config instead of failing.
• FR-4: Do all dollar arithmetic in the deterministic ranker (scripts/rank-and-report.py) — the LLM never eyeballs a number; every leak carries a kind of confirmed / estimated / at-risk.
• FR-5: Render the CFO report per references/cfo-output-format.md: a split headline that never sums confirmed and unconfirmed dollars under one verb, a trailing-30-day window stamp, a ranked table with a Confidence column, and a #1-line annualized callout — each fix a single config change.

Out of scope

• Authoring cost-control policy (cluster policies, spot configs) — that is the sibling databricks-cost-tuning skill; this skill detects and reports.
• Applying any fix — every remediation is a recommendation for a human-approved config change; the skill never mutates workspace configuration.
• Estimating spend when the billing tables are unreadable — no grant chain, no report (fail fast at Step 1 rather than substitute guesses).
• Standard-tier / non-Unity-Catalog workspaces — the system.* tables are UC-governed and unavailable there.