| name | learn-databricks |
| description | Databricks-specific lessons and knowledge base. 13 hard-won lessons from the REACH-ADHD deployment (VM stockout, init scripts, env vars, pandas compatibility, partition mismatches, MLflow signatures, Model Serving payload routing, CLI auth) plus the full deployment journey doc. Use when running pipelines on Databricks, debugging cluster/notebook issues, deploying to Model Serving, or onboarding. Verbs: lesson (capture/list/search gotchas), feedback (lesson doc gaps), digest (bulk harvest from session). Trigger: databricks, cluster issue, notebook, dbutils, pipeline on databricks, model serving, mlflow, /learn-databricks. |
| argument-hint | [lesson | feedback | digest] [args] |
| allowed-tools | Bash, Read, Write, Edit, Grep, Glob |
| metadata | {"version":"1.0.0","last_updated":"2026-06-26","summary":"Databricks lessons, journey docs, and feedback system."} |
Skill: learn-databricks
Databricks-specific knowledge base for the REACH team.
What It Contains
lesson/ 13 hard-won lessons from the REACH-ADHD Databricks deployment
(VM stockout, init scripts, env vars, pandas compat, partitions,
MLflow signatures, Model Serving payload routing, CLI auth)
feedbacks/ Journey docs + inbox for lesson gaps and errors
Existing Lessons
01 Azure VM stockout — use confidential compute (DC-series)
02 Init scripts blocked — use Libraries API
03 Use ML Runtime for ML packages
04 %pip magic only works interactively
05 dbutils.notebook.run() loses env vars
06 Use separate job tasks, not orchestrator notebook
07 Set all env vars explicitly per notebook
08 pandas version compatibility (1.5.3 vs 2.x)
09 Don't install many packages on small VMs
10 CaseSet partition mismatch — upload from local
11 Unity Catalog requires model signature
12 "dataframe_records" is reserved — use payload passthrough
13 Databricks CLI auth tokens expire (~1 hour)
14 Agent workspace on Databricks — use Claude Agent SDK
15 Policy-locked Standard (USER_ISOLATION) clusters forbid jobs — run inline
Verbs
/learn-databricks lesson "" Capture a new lesson
/learn-databricks lesson list List all lessons
/learn-databricks lesson search Find relevant lessons before acting
/learn-databricks feedback "" Report a lesson doc gap or error
/learn-databricks digest Bulk harvest feedback from a session
The Guardrail Contract
BEFORE executing Databricks pipeline or infrastructure work, the agent MUST:
- Scan lesson/ for relevant lessons
- Flag any that apply to the user
Entry Points
Before acting ..... grep lesson/*.md for relevant gotchas
After a surprise .. /learn-databricks lesson ""
Lesson is wrong ... /learn-databricks feedback "<what's wrong>"