A data platform organises tables into bronze, silver and gold layers. What determines what belongs in each, and what is the common failure?

What determines each layer

• Bronze: raw, as received, immutable, with provenance. No transformation beyond what is needed to land it. Its value is that it is reproducible — any downstream error can be corrected by reprocessing from bronze without re-fetching from the source, which may no longer have the data.
• Silver: cleaned, conformed, deduplicated, typed, with a defined schema. This is where quality rules and business keys are applied, and it is the layer most consumers should read from.
• Gold: aggregated and shaped for specific consumption — a dashboard, a metric, a machine-learning feature, a reverse-ETL destination.

The common failure

Layers as a naming convention rather than as a contract. Tables are placed in a layer by whoever created them, with no agreed meaning, so a "silver" table may be raw and a "gold" table may be a staging step.

The distinguishing property is the guarantee, not the position: silver guarantees a schema and a quality level; gold guarantees a definition. If a consumer cannot rely on that, the layers convey nothing.

The second common failure

Consumers reading from bronze because silver is missing something they need. That is a product signal — the silver layer does not serve them — and the response should be to extend silver rather than to accept the bypass, because a consumer on bronze is coupled to the source's raw schema and will break when the source changes.

What must be true regardless of layer count

• Reproducibility from the rawest layer, which is bronze's whole purpose.
• Lineage, so a wrong number in a dashboard can be traced to its inputs.
• Ownership per table, since a dataset with no owner is one nobody fixes.
• Compaction as a platform responsibility, because small-file proliferation from streaming writers slows every reader and no individual writer is responsible for it.

The honest note on layer count

Three is a convention, not a law. Some platforms need two, some need four, and the number matters far less than whether each layer has a stated guarantee that consumers can rely on.