Staging, Serving and Data Vault Layers

• Describe the bronze/silver/gold data flow
• Explain why raw must stay immutable
• Compare medallion layering with data-vault staging

• Immutable raw is the forensic record — any modeled table can be rebuilt from it.
• Layer separation gives each stage its own failure mode, owner and SLA.
• Governance can differ by layer: raw is locked, gold is curated and certified.
• When a 'silver' bug is found, you repair gold by building from raw, not by patching guesses.

Negatives are the untouched original (bronze), proofs are cleaned-up drafts (silver), and the final print arrives at the gallery (gold). If a print is bad, you never retouch a copy of a copy — you go back to the negative.

• Bronze/raw: source snapshots and events land verbatim, timestamped, write-once.
• Silver/curated: dedupe, typecast, conform entities (one customer id), enforce quality tests.
• Gold/serving: fact/dimension models, semantic views and certified aggregates for BI.
• Data vault layers stage hubs (entities), links (relationships) and satellites (attributes) for auditable lineage.
• Each layer has owned refresh jobs; lineage records the path from gold back to bronze.

• Every source lands in bronze as a full snapshot; silver builds one clean customer per source; gold models orders, revenue and churn for dashboards.
• A regulated bank keeps a data-vault style audit trail—every change to a customer attribute is a satellite row with valid_from/valid_to.
• When a gold aggregate is wrong, the team backfills from silver in minutes, then diff-reconciles.

• Treat bronze as append-only; never let a transform clobber a snapshot.
• Put quality gates between layers: silver should pass expectations before gold builds.
• Give each layer an owner and prefix (bronze_*, gold_*) so lineage is self-describing.
• Document the backfill procedure per layer — restore is a feature, not an afterthought.

• Letting reports read silver or bronze 'because it's faster' — definitions fork.
• Updating raw in place when a source re-emits, corrupting forensic history.
• Skipping silver and modeling from bronze repeatedly, spreading messy logic.
• Building four layers because the vendor sells four badges, not because you need them.

• Track the layer a discrepancy entered: diff bronze→silver→gold counts to localize.
• If gold is right but a silver table diverges, check the silver quality gates.
• Lineage from gold to bronze tells you which raw wrote the bad value.