Pipeline Cost Analyzer

Illustrative example — demonstrates a well-formed SKILL.md for this reference implementation. Adapt the content to your own org; do not deploy verbatim.

pipeline-cost-analyzer

Overview

Surfaces Databricks cost insights from the billing system tables: it ranks the most expensive jobs, tracks DBU consumption trends by SKU and product, and turns each result into concrete optimization actions. Scope is deliberately narrow — cost/usage analysis from system.billing.usage and system.lakeflow.jobs, nothing else.

When to use this skill

Reach for this skill whenever a user asks about Databricks spend, DBU consumption, or job-level cost attribution. Representative requests:

• "What are my top 10 most expensive jobs in the last 30 days?"
• "Show me DBU consumption by SKU over the last quarter."
• "Which jobs increased their DBU spend the most week-over-week?"
• "What is the cost trend for ALL_PURPOSE compute over the last 6 months?"
• "What is my total DBU spend this month vs. last month?"
• "Compare serverless vs. classic compute spend this quarter."
• "What drove the biggest cost spike last quarter?"

Instructions

When the user asks a cost or job-performance question:

1. Identify the time window (default: last 30 days if not specified).
2. Query system.billing.usage for DBU consumption.
3. Join with system.lakeflow.jobs when job-level attribution is needed.
4. Present results as a ranked table, then give 2–3 actionable recommendations (see the Recommendations framework below).

Examples

Top-spending jobs (last 30 days)

SELECT
  u.usage_metadata.job_id                     AS job_id,
  j.name                                       AS job_name,
  u.sku_name,
  SUM(u.usage_quantity)                        AS total_dbus
FROM system.billing.usage u
LEFT JOIN system.lakeflow.jobs j
  ON u.usage_metadata.job_id = j.job_id        -- both job_id columns are STRING
 AND u.workspace_id = j.workspace_id
WHERE u.usage_date >= DATE_SUB(CURRENT_DATE(), 30)
  AND u.usage_metadata.job_id IS NOT NULL
GROUP BY 1, 2, 3
ORDER BY total_dbus DESC
LIMIT 20

DBU spend by SKU — weekly trend (last 90 days)

SELECT
  sku_name,
  DATE_TRUNC('week', usage_date)               AS week,
  SUM(usage_quantity)                          AS total_dbus
FROM system.billing.usage
WHERE usage_date >= DATE_SUB(CURRENT_DATE(), 90)
GROUP BY 1, 2
ORDER BY 2, total_dbus DESC

Monthly cost trend by product (last 6 months)

SELECT
  DATE_TRUNC('month', usage_date)              AS month,
  billing_origin_product,
  SUM(usage_quantity)                          AS total_dbus
FROM system.billing.usage
WHERE usage_date >= DATE_SUB(CURRENT_DATE(), 180)
GROUP BY 1, 2
ORDER BY 1, total_dbus DESC

Recommendations framework

After presenting results, always include:

1. Top offender — the single highest-cost job or SKU and one concrete action (e.g. "set a max DBU budget on job X", "switch to job compute from all-purpose for batch jobs").
2. Trend signal — is spend rising, flat, or falling? Flag if up >20% month-over-month.
3. Quick win — one change achievable in under an hour (e.g. enable auto-termination on idle clusters, reduce cluster size on off-peak jobs).

Edge cases

• No job attribution — usage_metadata.job_id is NULL for non-job usage (all-purpose compute, SQL warehouses). Report those as unattributed rather than dropping them.
• Sparse windows — if the requested window has no usage rows, say so explicitly instead of returning an empty table with no explanation.
• STRING join keys — job_id and workspace_id are both STRING in these tables; never cast them to int for the join.

Data scope

• system.billing.usage — DBU consumption events per workspace/SKU/job/cluster
• system.lakeflow.jobs — job definitions (name, owner, schedule, tags)
• No PII or sensitive customer data is accessed.

These two tables are the skill's declared unity_catalog_scopes in registry.yaml — an optional data-footprint declaration. Runtime access is governed by Unity Catalog on whoever runs the skill; CI flags (advisory only) any query here that references a table outside the declared footprint.