e2e-medallion-architecture

Update Check — ONCE PER SESSION (mandatory) The first time this skill is used in a session, run the check-updates skill before proceeding.

• GitHub Copilot CLI / VS Code: invoke the check-updates skill.
• Claude Code / Cowork / Cursor / Windsurf / Codex: compare local vs remote package.json version.
• Skip if the check was already performed earlier in this session.

CRITICAL NOTES

1. To find the workspace details (including its ID) from workspace name: list all workspaces and, then, use JMESPath filtering
2. To find the item details (including its ID) from workspace ID, item type, and item name: list all items of that type in that workspace and, then, use JMESPath filtering

End-to-End Medallion Architecture

Prerequisite Knowledge

Read these companion documents — they contain the foundational context this skill depends on:

• COMMON-CORE.md — Fabric REST API patterns, authentication, token audiences, item discovery
• COMMON-CLI.md — az rest, az login, token acquisition, Fabric REST via CLI
• SPARK-AUTHORING-CORE.md — Notebook deployment, lakehouse creation, job execution
• notebook-api-operations.md — Required for notebook creation — .ipynb structure requirements, cell format, getDefinition/updateDefinition workflow

For Spark-specific optimization details, see data-engineering-patterns.md.

---

Architecture Overview

Medallion Architecture is a data lakehouse pattern with three progressive layers:

Layer	Purpose	Optimization Profile	Use Case
Bronze (Raw)	Land raw data exactly as received	Write-optimized, append-only, partitioned by ingestion date	Audit trail, reprocessing, lineage
Silver (Cleaned)	Deduplicated, validated, conformed data	Balanced read/write, partitioned by business date	Feature engineering, operational reporting
Gold (Aggregated)	Pre-calculated metrics for analytics	Read-optimized (ZORDER, compaction), partitioned by month/year	Power BI reports, dashboards, ad-hoc analytics via SQL endpoint

• Bronze: Schema-on-read — flexible schema, Delta time travel supports audit and rollback
• Silver: Schema enforcement — reject non-conforming writes; handle schema evolution with mergeSchema when sources change
• Gold: Strict schema governance — curated, business-approved datasets only

---

Must/Prefer/Avoid

MUST DO

• Choose lakehouse architecture based on schema-enabled availability (see infrastructure-orchestration.md):
  • Preferred: Schema-enabled lakehouse → create ONE workspace + ONE lakehouse with bronze, silver, gold schemas
  • Legacy: Non-schema-enabled → create separate workspaces per layer (Bronze, Silver, Gold) for governance and access control
• Use Livy API for schema and table creation — to create schemas and tables in a schema-enabled lakehouse, submit Spark SQL statements via Livy sessions (POST /livyApi/versions/2023-12-01/sessions → POST .../statements). This is the only programmatic REST path for DDL operations (CREATE SCHEMA, CREATE TABLE) in Fabric lakehouses.
• Add metadata columns in Bronze: ingestion timestamp, source file, batch ID
• Apply data quality rules in the Bronze-to-Silver transformation (deduplication, null handling, range validation)
• Use Delta Lake format for all medallion layer tables
• Use partition-aware overwrite in Silver/Gold writes to avoid reprocessing unchanged data
• Include validation steps after each layer (row counts, schema checks, anomaly detection)
• Follow the .ipynb validation + Fabric nuances in notebook-api-operations.md when creating notebooks via REST API — every code cell must include "outputs": [] and "execution_count": null
• Complete the full end-to-end flow — do not stop after creating notebooks; always bind lakehouses, execute notebooks sequentially (Bronze → Silver → Gold), verify results, and connect Power BI to the Gold layer unless the user explicitly requests a partial setup

PREFER

• Incremental processing (watermark pattern) over full refresh
• Separate notebooks per layer for independent testing and debugging
• ZORDER on frequently filtered columns in Gold tables
• Running OPTIMIZE after writes in Silver and Gold layers
• Environment-specific Spark configs (write-heavy for Bronze, balanced for Silver, read-heavy for Gold)
• OneLake shortcuts to expose Gold data to consumer workspaces without duplication
• Clear layer ownership: engineers own Bronze/Silver, analysts own Gold
• Fabric Variable Libraries to centralize paths and configuration across layers
• Multi-workspace deployment patterns for medium/high governance requirements (Bronze/Silver/Gold in separate workspaces)
• Use Materialized Lake Views (MLVs) for Silver/Gold tables when the transformation is expressible in Spark SQL and benefits from declarative refresh semantics. See spark-authoring-cli — Materialized Lake View patterns and MLV incremental refresh patterns.
• For MLV refresh scheduling and monitoring (create/delete schedules, trigger on-demand refresh, check job status), use mlv-operations-cli. Note: "materialized view", "spark materialized view", and "MLV" all refer to the same Fabric feature.

AVOID

• Storing all layers in a single lakehouse WITHOUT schemas — non-schema lakehouses require notebook init cells or Environment configuration to enable OneLake Spark Catalog for RLS/CLS and MLVs. Use separate lakehouses for isolation if schemas aren't available.
• Creating 3 separate lakehouses when schema-enabled lakehouse is available — use schemas within one lakehouse instead (cleaner, no boilerplate init cells, more efficient for MLV cross-schema transformations)
• Skipping the Silver layer and going directly from Bronze to Gold
• Hardcoded workspace IDs, lakehouse IDs, or FQDNs — discover via REST API
• SELECT * without LIMIT on Bronze tables (they grow unboundedly)
• Running VACUUM without checking downstream dependencies
• Chaining OneLake shortcuts between medallion layers (Bronze→Silver→Gold) — each layer must be physically materialized for lineage and governance
• Copying complete implementation code into skills — guide the LLM to generate instead
• Reading from external HTTP/HTTPS URLs directly in Spark — Fabric Spark cannot access arbitrary external URLs; land data in lakehouse Files/ first (via curl, OneLake API, or Fabric pipeline Copy activity), then read from the lakehouse path
• Creating notebooks via REST API without validating .ipynb structure — missing execution_count: null or outputs: [] on code cells causes silent failures or "Job instance failed without detail error"

---

Workspace Setup Guidance

When setting up a medallion workspace, choose your architecture pattern first (see infrastructure-orchestration.md for detailed guidance):

Option A: Schema-Enabled Lakehouse (Preferred)

1. Create single workspace: {project}-{env}
2. Create one lakehouse with schemas: {project}_lakehouse
3. Create schemas within the lakehouse:
   • bronze schema for raw ingestion
   • silver schema for cleaned/validated data
   • gold schema for aggregated analytics
4. Choose transformation approach:
   • Option 4a: Use notebooks for each layer (PySpark or Spark SQL transformations)
   • Option 4b: Use Materialized Lake Views (Spark SQL) for declarative transformations with incremental refresh (when query is IR-eligible) — see materialized-lake-view-patterns.md and mlv-incremental-refresh-patterns.md
   • Note: PySpark MLVs exist but use full refresh only (no incremental) — use when you need UDFs/complex Python logic
   • MLV benefit: OneLake Spark Catalog is automatically enabled for schema-enabled lakehouses — MLVs work out-of-box with no notebook init cells or Environment configuration required
5. RBAC (optional): Use row-level security and column masking within schemas for fine-grained access control (also requires OneLake Spark Catalog)

Option B: Separate Lakehouses (Legacy)

1. Create three workspaces:
   • {project}-bronze-{env}
   • {project}-silver-{env}
   • {project}-gold-{env}
2. Create one lakehouse per workspace:
   • Bronze workspace → {project}_bronze lakehouse
   • Silver workspace → {project}_silver lakehouse
   • Gold workspace → {project}_gold lakehouse
3. Assign RBAC per layer workspace:
   • Bronze: ingestion/engineering write permissions
   • Silver: engineering/data quality permissions
   • Gold: analytics/BI consumer access with stricter curation controls
4. Enable OneLake Spark Catalog for non-schema lakehouses (required for RLS/CLS and catalog-backed access patterns):
   • Primary: Set spark.sql.fabric.catalog.enable-schemaless-lakehouses=true in an Environment and attach it to notebooks.
   • Alternative: Omit default lakehouse binding from notebooks. Use four-part fully-qualified references (workspace.lakehouse.schema.table). OneLake Spark Catalog auto-enables when no default lakehouse is set.
   • Alternative (internal/unsupported): Add this as the first cell in every notebook:

   %%pyspark
   !echo "spark.sql.fabric.catalog.enable-schemaless-lakehouses=true" >> /home/trusted-service-user/.trident-context
   

   • ⚠️ Note: This workaround uses an internal runtime configuration path that may change in future Fabric releases. Prefer schema-enabled lakehouses for stable, documented OneLake Spark Catalog support.
   • With this configuration, non-schema lakehouses support:
     • ✅ Row-level security (RLS) and column-level security (CLS)
   • Note: MLVs require schema-enabled lakehouses (Option A). For non-schema lakehouses, use notebooks with Delta tables.

Common Steps (Both Options)

After completing Option A or Option B above, perform these steps:

1. Create notebooks for each layer (one per transformation stage) — follow .ipynb validation + Fabric nuances
2. Bind each notebook to its lakehouse — set metadata.dependencies.lakehouse with the correct lakehouse ID (see notebook-api-operations.md § Default Lakehouse Binding):
   • Option A: All notebooks → same lakehouse, use schema prefixes (bronze.table, silver.table)
   • Option B:
     • Bronze notebook → Bronze workspace/lakehouse
     • Silver notebook → Silver workspace/lakehouse (reads Bronze via cross-workspace OneLake access / fully qualified references)
     • Gold notebook → Gold workspace/lakehouse (reads Silver via cross-workspace access)
3. Confirm notebook deployment — check that updateDefinition returned Succeeded; this is sufficient confirmation that content and lakehouse binding persisted. Do NOT call getDefinition to re-verify — it is an async LRO and adds unnecessary latency.
4. Execute notebooks sequentially — Bronze first, then Silver, then Gold — using POST .../jobs/instances?jobType=RunNotebook with the correct defaultLakehouse in execution config (both id and name required)
5. Connect Power BI to Gold layer — discover the Gold lakehouse SQL endpoint, create a Direct Lake semantic model, create a report with visuals on the Gold summary table (see Gold Layer → Power BI Consumption)
6. Create pipeline to orchestrate the Bronze → Silver → Gold flow for recurring execution

Explicit Override: Single Workspace

If the user explicitly asks for a single workspace deployment (for example, POC/small team/monolithic pattern), keep the current approach:

• One workspace with separate Bronze/Silver/Gold lakehouses
• Preserve layer separation logically even when workspace is shared
• Call out governance trade-offs versus multi-workspace design

Parameterize by environment: workspace name suffix (-dev, -prod), data volume (sample vs full), capacity SKU, and Bronze retention period.

---

Bronze Layer — Ingestion Patterns

When a user requests data ingestion into the Bronze layer, guide LLM to:

1. Land data in lakehouse first: External data must be staged into the lakehouse Files/ folder before Spark can read it — use one of:
   • Fabric Pipeline Copy activity (preferred for recurring loads) — connects to external sources (HTTP, FTP, databases, cloud storage) and writes to OneLake
   • OneLake API / curl — upload files via REST API using storage.azure.com token (see COMMON-CLI.md § OneLake Data Access)
   • OneLake Shortcut — for data already in Azure ADLS Gen2, S3, or another OneLake location
   • notebookutils.fs — copy from mounted storage paths within a notebook
   • ⚠️ Fabric Spark cannot read from arbitrary HTTP/HTTPS URLs — spark.read.format("csv").load("https://...") will fail
2. Read from lakehouse path: Once data is in Files/, read using lakehouse-relative paths (e.g., spark.read.format("csv").load("Files/landing/daily/"))
3. Add metadata and write: Tracking columns (ingestion timestamp, source file, batch ID), Delta table with descriptive name, partition by ingestion date, append mode
4. Validate: Log row counts, validate schema structure, flag anomalies vs historical patterns

---

Silver Layer — Transformation Patterns

When a user requests Bronze-to-Silver transformation, guide LLM to:

• Quality rules: Deduplicate on natural/composite key, filter invalid ranges, handle nulls (drop required, fill optional), validate logical constraints
• Schema conformance: snake_case column names, standardized data types, derived columns (durations, percentages, categories)
• Schema evolution: Use mergeSchema option when source schemas change; coordinate downstream updates to Gold tables and Power BI datasets
• Write strategy: Partition by business date, partition-aware overwrite, run OPTIMIZE after write, log before/after metrics

---

Gold Layer — Aggregation Patterns

When a user requests Gold analytics tables, guide LLM to generate:

• Common aggregates: Daily/weekly/monthly summaries, dimensional analysis (by location, category, type), trend breakdowns over time, demand patterns (hour-of-day, day-of-week)
• Spark session config — set these properties in the Gold notebook before any write operations:

  spark.conf.set("spark.sql.parquet.vorder.default", "true")
  spark.conf.set("spark.databricks.delta.optimizeWrite.enabled", "true")
  spark.conf.set("spark.databricks.delta.optimizeWrite.binSize", "1g")
  

  • V-Order (vorder.default) — applies Fabric's columnar sort optimization to all Parquet files, dramatically improving Direct Lake and SQL endpoint read performance
  • Optimize Write (optimizeWrite.enabled) — coalesces small partitions into optimally-sized files (target ~1 GB per binSize), reducing file count and improving scan efficiency
• Optimization: ZORDER on filter columns, run OPTIMIZE after writes, pre-aggregate metrics to avoid runtime computation

---

End-to-End Execution Flow

When setting up medallion architecture end-to-end, the LLM must not stop after creating notebooks and deploying code. The complete lifecycle is:

Create Resources → Deploy Content → Bind Lakehouses → Execute → Verify Results

Step-by-Step

1. Create layer workspaces and lakehouses (default) — one workspace and one lakehouse per layer (Bronze, Silver, Gold); capture workspace IDs and lakehouse IDs
2. Create notebooks — one per layer, with valid .ipynb structure (see notebook-api-operations.md)
3. Bind lakehouse to each notebook — include metadata.dependencies.lakehouse in the .ipynb payload with:
   • default_lakehouse: the target lakehouse GUID
   • default_lakehouse_name: the lakehouse display name
   • default_lakehouse_workspace_id: the workspace GUID
4. Deploy notebook content — updateDefinition with the Base64-encoded .ipynb payload (content + lakehouse binding together)
5. Confirm deployment — check that each updateDefinition LRO returned Succeeded; that is sufficient. Do NOT call getDefinition to re-verify — it is an async LRO and adds significant latency per notebook.
6. Execute notebooks sequentially — use POST .../jobs/instances?jobType=RunNotebook:
   • Pass defaultLakehouse with both id and name in executionData.configuration
   • Run Bronze first → poll until Completed → run Silver → poll → run Gold → poll
   • Check for recent jobs before submitting (prevent duplicates — see SPARK-AUTHORING-CORE.md)
7. Verify results — after each notebook completes, confirm expected tables exist and row counts are reasonable
8. Connect Power BI to Gold — create semantic model + report on Gold summary tables (see Gold Layer → Power BI Consumption)

Common Failure: Stopping After Notebook Creation

If the flow stops after deploying notebook code without binding or executing:

• Notebooks will have no lakehouse context → spark.sql() and relative paths (Tables/, Files/) fail at runtime
• The user sees no output or results — the architecture is set up but never tested
• Always complete through step 7 unless the user explicitly asks to stop at a specific step

---

Gold Layer → Power BI Consumption

After Gold tables are populated, connect Power BI to surface the analytics. Build a semantic model on top of the Gold lakehouse, using DirectLake.

Step-by-Step

1. Discover the Gold lakehouse SQL endpoint — call GET /v1/workspaces/{workspaceId}/lakehouses/{goldLakehouseId} and extract properties.sqlEndpointProperties.connectionString and provisioningStatus; wait until status is Success
2. Verify Gold tables via SQL — connect to the SQL endpoint using sqlcmd (see COMMON-CLI.md § SQL / TDS Data-Plane Access) and confirm the target table exists:

   SELECT TABLE_NAME FROM INFORMATION_SCHEMA.TABLES WHERE TABLE_NAME = 'nyc_taxi_daily_summary'
   

3. Create a semantic model — use the semantic-model-authoring skill for semantic model creation and TMDL deployment. Create via POST /v1/workspaces/{workspaceId}/items with type: "SemanticModel" then deploy definition via updateDefinition using TMDL format (see ITEM-DEFINITIONS-CORE.md § SemanticModel):
   • The model must reference the Gold lakehouse SQL endpoint as its data source
   • Define a table mapping to the Gold summary table (e.g., nyc_taxi_daily_summary)
   • Use Direct Lake mode — this connects directly to Delta tables in OneLake without data import
   • Include measures for key aggregations you find interesting (e.g., Total Trips, Avg Fare, Total Revenue, Month over Month Growth)
4. Create a Power BI report — POST /v1/workspaces/{workspaceId}/items with type: "Report" then deploy definition via updateDefinition using PBIR format (see ITEM-DEFINITIONS-CORE.md § Report):
   • Reference the semantic model created in step 3 via definition.pbir
   • Define at least one page with visuals on the Gold summary table
   • Suggested visuals: line chart (daily trend), card (KPI totals), bar chart (by category), table (detail view)
5. Verify end-to-end — use the semantic-model-consumption skill to run DAX queries against the semantic model and confirm data flows from Gold tables through to the report

Principles

• Discover SQL endpoint dynamically — the connection string is in properties.sqlEndpointProperties.connectionString on the lakehouse response; never hardcode it
• Wait for SQL endpoint provisioning — status must be Success before connecting; newly created lakehouses may take minutes to provision
• Prefer Direct Lake mode — avoids data duplication; semantic model reads directly from OneLake Delta tables
• Match table/column names exactly — the semantic model table definition must use the exact Delta table and column names from the Gold lakehouse
• For semantic model authoring (TMDL, refresh, permissions), cross-reference the semantic-model-authoring skill
• For DAX query validation, cross-reference the semantic-model-consumption skill

---

Pipeline Orchestration

When a user requests a pipeline for the medallion flow, guide LLM to design with:

• Structure: Sequential activities (Bronze → Silver → Gold), each waiting for previous success; independent Gold aggregations can run in parallel; include validation and notification activities
• Parameterization: Pipeline-level processing date (defaults to yesterday), passed to all notebooks; dynamic date expressions
• Scheduling: Daily aligned with source refresh, watermark-based incremental processing, periodic full refresh for corrections
• Error handling: Retry with backoff for transient failures, alerting for persistent failures, graceful degradation (downstream uses previous data if upstream fails)

---

Environment Optimization

For detailed Spark configurations and optimization strategies, see data-engineering-patterns.md.

Layer	Profile	Key Settings
Bronze	Write-heavy	Disable V-Order, enable autoCompact, large file targets, partition by ingestion_date
Silver	Balanced	Enable V-Order, adaptive query execution, partition by business date, ZORDER on filtered columns
Gold	Read-heavy	V-Order (spark.sql.parquet.vorder.default=true), Optimize Write (optimizeWrite.enabled=true, binSize=1g), vectorized readers, adaptive execution, ZORDER on all filter columns, pre-aggregate metrics

---

Examples

Example 1: Set Up Medallion Workspaces (Default)

Prompt: "Set up medallion architecture with separate Bronze, Silver, and Gold workspaces for sales analytics"

What the LLM should generate: REST API calls to:

1. Create workspaces: sales-bronze-dev, sales-silver-dev, sales-gold-dev
2. Create one lakehouse in each workspace: sales_bronze, sales_silver, sales_gold
3. Assign RBAC roles per workspace/layer

# Workspace creation (see COMMON-CLI.md for full patterns)
cat > /tmp/body.json << 'EOF'
{"displayName": "sales-analytics-dev"}
EOF
workspace_id=$(az rest --method post --resource "https://api.fabric.microsoft.com" \
  --url "https://api.fabric.microsoft.com/v1/workspaces" \
  --body @/tmp/body.json --query "id" --output tsv)

# Create Bronze lakehouse
cat > /tmp/body.json << 'EOF'
{"displayName": "sales_bronze", "type": "Lakehouse"}
EOF
az rest --method post --resource "https://api.fabric.microsoft.com" \
  --url "https://api.fabric.microsoft.com/v1/workspaces/$workspace_id/items" \
  --body @/tmp/body.json

Example 2: Design Bronze Ingestion

Prompt: "Ingest daily CSV files into bronze lakehouse with metadata columns"

What the LLM should generate: PySpark notebook that:

1. Reads source files with schema inference or explicit schema
2. Adds ingestion_timestamp, source_file, batch_id columns
3. Writes to Delta table partitioned by ingestion date
4. Logs row count and validation metrics

# Bronze ingestion pattern (guide LLM to generate full implementation)
from pyspark.sql.functions import current_timestamp, input_file_name, lit
import uuid

batch_id = str(uuid.uuid4())
df = (spark.read.format("csv").option("header", True).load("/Files/landing/daily/")
      .withColumn("ingestion_timestamp", current_timestamp())
      .withColumn("source_file", input_file_name())
      .withColumn("batch_id", lit(batch_id)))
df.write.mode("append").partitionBy("ingestion_date").format("delta").saveAsTable("bronze.events_raw")

Example 3: Bronze-to-Silver Transformation

Prompt: "Clean bronze data: remove duplicates, filter invalid records, add derived columns, write to silver"

What the LLM should generate: PySpark notebook applying quality rules, schema conformance, and partitioned write with optimization.

Example 4: End-to-End Pipeline

Prompt: "Create a pipeline that runs bronze ingestion, then silver transformation, then gold aggregation daily at 2 AM"

What the LLM should generate: Pipeline JSON definition with sequential notebook activities, date parameter, retry logic, and schedule trigger.