Data Pipeline Spec

Data Pipeline Spec Skill

This skill produces a complete data pipeline specification covering sources, transformations, destinations, scheduling, SLAs, error handling, data quality checks, and monitoring requirements. Output is ready for engineering handoff or architecture review.

Required Inputs

Ask the user for these if not provided:

• Pipeline purpose — what business question or workflow does this pipeline serve?
• Source systems — where does data come from? (databases, APIs, files, event streams)
• Destination — where does data land? (data warehouse, data lake, downstream DB, reporting tool)
• Transformation type — ETL (transform before loading) or ELT (load raw, transform in warehouse)?
• Frequency / SLA — how often must data be fresh? (real-time / hourly / daily / weekly)
• Volume estimate — approximate rows/events per run
• Data quality requirements — completeness, deduplication, freshness, schema enforcement
• Team or stack — any specific tools in use? (Airflow, dbt, Fivetran, Spark, Kafka, etc.)

Output Structure

Data Pipeline Spec: [Pipeline Name]

Purpose: [One sentence — what decision or workflow does this pipeline enable?] Type: [ETL / ELT / Streaming / Batch] Owner: [Team or individual] Version: [1.0] Date: [Date] Status: [Draft / Under Review / Approved]

1. Overview

[2–3 sentences describing the pipeline end-to-end: what data moves, from where to where, at what cadence, and why.]

Architecture diagram (text):

[Source A] ──┐
[Source B] ──┤──► [Ingestion Layer] ──► [Transform Layer] ──► [Destination] ──► [Consumers]
[Source C] ──┘

2. Sources

Incremental key (if applicable): [The column used to identify new or changed records — e.g. updated_at, event_id]

Authentication: [API key / OAuth / IAM role / connection string — note where credentials are stored]

3. Ingestion Layer

Tool: [Fivetran / Airbyte / Kafka Connect / custom script / dbt source]

Ingestion method:

• Full extract (full table refresh each run)
• Incremental extract (only new/changed rows since last run)
• CDC (change data capture from database transaction log)
• Event streaming (continuous ingestion from Kafka/Kinesis)

Raw landing zone: [Where raw data lands before transformation — e.g. raw.salesforce_opportunities in Snowflake, S3 bucket s3://data-raw/crm/]

Schema handling: [Strict schema enforcement / Schema evolution allowed / Union schema]

4. Transformation Logic

List each transformation in execution order. For ELT pipelines, this is the dbt model or SQL layer.

Business logic rules:

• [e.g. Revenue is recognised on payment_confirmed_at, not payment_initiated_at]
• [e.g. Users in the internal@company.com domain are excluded from all metrics]
• [e.g. Currency conversion uses the ECB rate from the first business day of each month]

Slowly Changing Dimensions (SCD) — if applicable:

• [e.g. users.plan_tier is SCD Type 2 — keep history of plan changes with valid_from / valid_to]

5. Destination

Partitioning / Clustering: [e.g. Partitioned by event_date, clustered by user_id — reduces query cost for time-range scans]

Retention policy: [e.g. Raw data retained for 90 days; mart tables retained indefinitely]

6. Scheduling & SLAs

Schedule: [Cron expression and human description — e.g. 0 6 * * * — daily at 06:00 UTC]

Trigger type:

• Time-based (cron)
• Event-based (triggered by upstream pipeline success / file arrival / Kafka lag)
• Manual (ad hoc runs only)

Backfill strategy: [How to reprocess historical data if the pipeline fails or logic changes — e.g. parameterised date range, full drop-and-reload]

7. Data Quality Rules

DQ tool: [dbt tests / Great Expectations / Monte Carlo / custom SQL assertions]