System design interview

Medallion Architecture (Bronze, Silver, Gold)

MediumFree75 min read

Design a lakehouse with Bronze (raw), Silver (cleaned), and Gold (business aggregates), including tier boundaries and consumers.

system-designmedallionlakehousearchitecture-patterns

Start here

The problem in plain words

An analyst builds a dashboard straight on top of raw files in the data lake, for example every JSON file under a folder called orders. The daily revenue chart works, and it becomes the executive favourite.

Then a team renames a field from order_total to order_amount during a code clean-up. The files still arrive, the pipeline still succeeds, and the dashboard quietly reads nothing for that field. Revenue shows zero for two days. No alarm rang, and leadership stops trusting the lake.

The medallion design is a way to stop this. It sorts data into three layers with clear rules: Bronze keeps the raw evidence, Silver holds cleaned and checked data, and Gold holds the business numbers people actually read. Each layer has an owner and a promise.

Words you will meet

Data lake
Cheap storage holding raw files of all kinds, for example in Amazon S3.
Lakehouse
A data lake with a table layer on top, so you can query the files like database tables and get reliable updates.
Bronze
The landing zone. Raw data exactly as it arrived, only ever added to, with a note of where and when it came from.
Silver
Data that is parsed, typed, de-duplicated and checked. The trusted place to join things.
Gold
Ready-made business tables and metrics, such as daily revenue, built for dashboards.
Append-only
You only add new data. You never overwrite or delete the original.
Data contract
An agreement between a producer and a consumer about field names, types and meaning, with a check that fails loudly when it is broken.
Quarantine
A side table for rows that failed checks, so you can inspect them without losing them or polluting good data.
Schema evolution
A table's columns changing over time, such as a field being added or renamed.
A dashboard on raw files, and what a rename does to it
renamereadszeroOrder servicerenames a fieldRaw JSON filesorder_total is noworder_amountRevenue dashboardreads order_totalLeadershipsees zero revenue
Press Play to watch it move

Each arrow is one thing happening. Play it step by step, or read the list below.

  1. 1.A service team renames order_total to order_amount. Files keep landing and nothing fails.
  2. 2.The dashboard reads raw files directly, looks for order_total and finds nothing.
  3. 3.Revenue shows as zero for two days. There was no failed pipeline and no alert, so trust in the lake erodes.

The numbers, as things you can picture

  • Many source systemsEach system sends data in its own shape, so one shared place with rules is needed.
  • Daily batches and streamingBoth kinds of input land in the same Bronze layer.

What the interviewer wants

Define what Bronze, Silver and Gold are for, what contracts and checks sit between them, and who is allowed to read each layer, including how you fix a mistake by replaying from Bronze.

Interview framing

You are designing a medallion lakehouse in a mid-level DE interview. The interviewer wants Bronze (raw), Silver (cleaned), and Gold (business aggregates), plus clear boundaries and consumer rules.

They are testing whether you understand that a lake without contracts becomes a swamp, and whether you can explain replay, quality gates, and who is allowed to query what.

Background from first principles

A data lake is cheap object storage filled with files. Anyone can dump JSON. Anyone can query it with a Spark job. That flexibility is useful. It is also how companies destroy trust.

Without layers and ownership:

  • Analysts build dashboards on raw dumps.
  • A producer renames a field.
  • Revenue silently zeros out.
  • Leadership stops believing "the lake."

Medallion architecture is a social and technical contract. Raw stays raw so you can recover. Cleaned data is curated with schemas and keys. Business numbers live in Gold with owners, tests, and stable names.

Bronze, Silver, and Gold are not magic Databricks trademarks you must invoke. They are 3 jobs for data at different distances from source truth and business meaning.

Analogy

Think of a professional kitchen.

  • Bronze is the delivery dock: crates arrive labeled, you do not cook yet, you keep the packing slips.
  • Silver is mise en place: washed, cut, typed ingredients ready for recipes.
  • Gold is the plated dish: intentional, consistent, what the diner (BI, exec dashboard) should see.

If diners wander into the dock and cook from crates, food poisoning is on you.

Full problem statement

Design a lakehouse with:

  • Bronze: raw, append-only, recoverable.
  • Silver: parsed, typed, deduplicated, conformed entities or events.
  • Gold: business marts and KPIs for BI and product metrics.

Define:

  • Boundaries between tiers.
  • Who may read each tier.
  • Quality gates between Silver and Gold.
  • How replay works when Silver logic changes.
  • How schema evolution and PII rules apply.

Context: multiple source systems, daily and streaming inputs. BI users should not query Bronze for production KPIs.

What "good" looks like in the room

You open with a trust-failure story. You define each tier with grain and responsibilities. You draw the flow. You state access rules. You explain replay from Bronze. You name quality checks at boundaries. You discuss whether Silver is entity-current or immutable events. You keep Gold contracts stable.

Clarifying questions

  • Streaming into Bronze, batch dumps, or both?
  • Are Silver tables entity-current (SCD-ish) or immutable event facts?
  • Do data scientists get Silver access for exploration and training?
  • Tooling: Delta/Iceberg + dbt, or something else already standard here?
  • Which domains own which Gold marts?
  • What is the freshness SLA for Gold executive dashboards?

Scale and estimation prompts

Even without hyperscale numbers, show you think in volumes:

  • How many source systems and daily file volume?
  • Partition Bronze by ingest date for prune-friendly replay.
  • Estimate rebuild cost: reprocessing 30 days of Bronze into Silver when a dedupe key changes.
  • Concurrent BI queries should hit Gold, not scan raw JSON.

What is out of scope

  • Building a full data mesh org chart unless asked.
  • Vendor-only answers ("we use product X so medallion is done").
  • Letting every dashboard author define Gold metrics in a BI tool with no versioning.
  • Treating Bronze as a place for business cleanses that throw away evidence.

Expanded problem narrative

You are the first platform DE at a mid-size retailer. Sources include:

  • Orders OLTP database (nightly dump + CDC).
  • Clickstream Kafka topic.
  • SaaS marketing export (CSV daily).
  • Customer service tickets (API pull).

Executives want a single "trusted revenue" number. Data scientists want historical click features. Analysts want to explore quickly. If you give everyone Bronze access, you recreate the swamp. If you over-gate Silver so nobody can explore, teams build shadow pipelines in Sheets.

Your design must make the happy path obvious: producers land in Bronze, engineering publishes Silver with keys and types, analytics engineering publishes Gold marts with tests, BI reads Gold.

What the interviewer probes next

They will ask how you debug a wrong Gold metric upward. They will ask whether ML may read Silver. They will ask how you stop dashboards from scanning raw JSON. They will ask where masking applies. Have crisp answers ready.

Room diagram expectation

They expect something like:

Talk while you draw. Silence with boxes is weaker than a spoken walkthrough of one order id.