Applies toLakehouses and warehouses on Databricks, Snowflake, Fabric, Synapse, Redshift, BigQuery; supporting orchestration, ingestion and BI
Scope and preparation
| Step | Detail |
|---|---|
| Scope | One platform or domain; list in-scope workspaces/accounts, pipelines, BI |
| People | Platform lead, data engineers, security, a key consumer, finance/FinOps |
| Evidence | Architecture diagram, cost reports, incident history, access model, SLOs, runbooks |
| Framework | Cloud provider well-architected guidance plus the data-specific questions below |
Question set
Reliability
- Which datasets have SLOs, and how often are they missed?
- Can pipelines be re-run idempotently and backfilled safely?
- Is there a tested DR plan for storage, metadata/catalog and orchestration?
- Are checkpoints, state and table history protected?
Security and governance
- Is access granted to groups through a catalog, with least privilege?
- Are sensitive columns classified, masked or row-filtered?
- Is network access private? Are public endpoints disabled?
- Are secrets in a vault, and do production jobs run as service identities?
- Is audit logging enabled and reviewed?
Cost
- Is cost visible per team (tags/showback)?
- Do clusters/warehouses auto-stop? Is idle compute tracked?
- Are storage history and retention tuned?
- Is there a regular optimisation review?
Performance
- Are large tables clustered/compacted and statistics maintained?
- Are slow queries and jobs tracked with owners?
- Is compute sized from evidence, not guesswork?
Operational excellence
- Is everything deployed through CI/CD and infrastructure as code?
- Are alerts routed to owners with runbooks?
- Is there an inventory of pipelines and datasets with owners?
- Are platform upgrades (runtimes, connectors) on a schedule?
Data quality and lifecycle
- Are critical datasets tested and monitored?
- Do contracts exist for key interfaces?
- Are retention and deletion automated?
Rate and prioritise
| Rating | Meaning |
|---|---|
| High risk | Could cause major outage, breach or large cost; fix within weeks |
| Medium | Notable gap; plan within the quarter |
| Low | Improvement; backlog |
Keep the action plan to the top 10 items, each with an owner and target date. Re-review in 6–12 months.
Common findings
- Personal accounts owning production jobs.
- No DR test for the catalog/metastore or orchestration.
- Shared all-purpose clusters running production jobs.
- Raw data kept forever, including personal data.
- No SLOs, so the team learns about late data from users.
Tip
Send the question set ahead and ask for evidence links. The workshop then spends time on decisions, not on finding diagrams.
Checklist
- Scope, people and evidence agreed
- Question set covered across all pillars
- Risks rated
- Top 10 actions with owners and dates
- Follow-up review scheduled