lakehouse-pipeline-design（湖屋管道设计）

Lakehouse pipeline design (Databricks)

Use this skill when someone asks for a pipeline design, DLT design, ETL plan, CDC ingestion, or a review of an existing pipeline.

Deliverables

When activated, produce at least:

1. A filled design doc based on assets/pipeline-design-doc.md
2. A short, actionable implementation checklist (you can reuse references/pipeline-checklist.md)

Optionally (only if asked): a code skeleton (PySpark / SQL / DLT) that matches the design.

Minimal inputs (ask only what’s missing)

Ask up to 3 questions total. Prefer defaults.

• Source type: files / DB / API / Kafka / etc.
• Mode: batch / streaming / CDC
• Target: tables (catalog.schema.*) and consumers (dashboards, ML, downstream jobs)
• Volume + SLA: rows/day, latency/freshness SLO, cost constraints
• Governance: PII? UC catalogs/schemas, access groups

Design guidance (what to include)

• Architecture: bronze → silver → gold; DLT vs Jobs; where to enforce quality
• Incremental strategy: watermarking, MERGE for CDC, idempotency
• Delta table design: partitioning, ZORDER, OPTIMIZE/VACUUM policy
• Quality checks: schema validation, null/unique, freshness, anomaly checks
• Observability: metrics, logs, expectations failures, alerts, runbooks
• Backfills: replay strategy, how to reprocess safely, versioning
• Security: UC permissions, row/column filtering if needed, secrets management
• Operational: retries, SLAs, escalation, deployment strategy

Output rules

• Put concrete decisions in a “Decisions” section and unknowns in “Open questions”.
• If details are missing, keep placeholders like {{...}} and add an “Info needed” section.
• Keep the doc concise; link to references/pipeline-checklist.md when you need long checklists.

Examples

User: “Design a DLT pipeline that ingests Salesforce accounts daily and publishes a gold table for dashboards.” Output: Design doc + checklist + optional DLT skeleton.

User: “Review our existing silver-to-gold job for performance and reliability.” Output: Review-style design doc: risks, improvements, and prioritized actions.

Edge cases

• Streaming sources: include checkpointing, schema evolution handling, and late data policy.
• Regulated data: include classification, retention, and UC policy controls.
• Multi-tenant tables: call out tenant key, partitioning, and access controls.