synapse-migration

Telemetry — MANDATORY. Every api.fabric.microsoft.com call must carry x-ms-fabric-skill: synapse-migration (az rest: --headers "x-ms-fabric-skill=synapse-migration"), including every LRO poll, fabric_lro and retry. Snippets omit it — add it anyway.

CRITICAL NOTES

1. To find workspace details (including its ID) from a workspace name: list all workspaces, then use JMESPath filtering
2. To find item details (including its ID) from workspace ID, item type, and item name: list all items of that type in that workspace, then use JMESPath filtering
3. mssparkutils and notebookutils share the same API surface in most cases — the namespace is the primary change
4. Linked Services have no direct REST API equivalent in Fabric — they are replaced by Data Connections (for external sources) and OneLake Shortcuts (for storage mounts)
5. The Dedicated SQL Pool-to-Fabric-Lakehouse path is source- and feature-driven. You MUST complete these phases in strict order:
   1. Gap Assessment (MANDATORY FIRST): Run compatibility assessment using dedicated-pool-gap-assessment.md to identify unsupported features, blockers, and migration risks BEFORE attempting any conversion.
   2. User Approval (REQUIRED): After assessment, present 1:1, N:1, and N:N stored-procedure-to-notebook mapping strategies and require the user to explicitly approve the complete mapping, target names, dependency grouping, and workspace placement. Wait for explicit approval before proceeding.
   3. Manifest Creation (REQUIRED): Create migration-manifest.json to track all objects, approved mappings, conversion status, and deployment checkpoints. Update manifest atomically after every phase.
   4. Conversion: Generate artifacts only after approval is recorded in manifest. The 1:1, N:1, and N:N mapping applies ONLY to stored procedures, NOT to views. Views are schema objects like tables and are deployed via Livy as SQL view definitions; they never convert to notebooks. Preserve every procedure as an independently traceable source decision under any approved strategy. Stored-procedure transformation logic must use readable Spark SQL %%sql cells, not PySpark or the DataFrame API. Generated notebooks are outputs, not orchestration dependencies.
6. Treat the source Dedicated Pool as read-only. Never create sample or synthetic data, schemas, tables, views, procedures, users, roles, or grants in the source. Do not run source DDL, DML, or stored procedures. Source operations in this workflow are limited to metadata discovery and metadata-based validation; route any requested source mutation to a separate, explicitly scoped workflow.
7. Dedicated Pool data migration is out of scope. Never export, stage, copy, upload, shortcut, transfer, or load source table rows, and never run row-level or business-result equivalence queries. This skill migrates schema and executable code artifacts only.
8. Print regular migration status. Announce every step before it starts, report each object or checkpoint as it completes or fails, emit a concise heartbeat at least every 30-60 seconds during long-running operations, and close every phase with completed/failed/skipped/pending counts. Include the phase, step, object, state, elapsed time, and next action. Status output must exclude credentials, tokens, connection strings, and sensitive data values.
9. For large or complex stored procedures, prove conversion coverage with a deterministic source-block ledger and publish only from a hash-verified immutable deployment package. Every block must be converted, explicitly excluded with approval, or manually reviewed and approved; notebook syntax success alone is insufficient.
10. Before generating Dedicated Pool artifacts, read both resources/dedicated-pool-to-lakehouse.md and resources/dedicated-pool-conversion.md. In manifest parameterMappings, use sourceParameter for the discovered source name; parameterName is reserved for object-valued conf entries inside notebook %%configure.
11. A local or offline Dedicated Pool artifact request still requires this skill. For a large or complex procedure, also read resources/dedicated-pool-large-procedure-audit.md before generating artifacts.

Synapse Analytics → Microsoft Fabric Migration

Prerequisite Knowledge

These companion documents provide general Fabric REST patterns. Do NOT read them upfront — reference only when a specific phase requires a pattern not already covered in this skill's resource files:

• COMMON-CORE.md — General Fabric REST API patterns, authentication & token audiences, item discovery via JMESPath
• COMMON-CLI.md — az rest / az login CLI patterns, authentication recipes
• SPARK-AUTHORING-CORE.md — Notebook/lakehouse creation (already covered in spark-item-migration.md and lake-database-migration.md)
• SQLDW-AUTHORING-CORE.md — Fabric Warehouse T-SQL (delegate to sqldw-cli skill)

Auth, API endpoints, and item payloads are fully documented in this skill's own files. The common docs above are fallback references only.

---

Table of Contents

Context Loading Guide

IMPORTANT — Load only what you need. Do NOT read all resource files upfront. Load the specific file for the phase you are executing:

---

API-Driven Migration Workflow

This skill supports programmatic migration of Synapse Spark items via REST APIs (no UI-based Migration Assistant required).

Authentication

Use the token-acquisition recipe in COMMON-CLI § Authentication Recipes with the audiences above.

Migration Phases (Execute in Order)

Phase order matters: Environments (Phase 0) must exist before notebooks/SJDs can bind to them. Lakehouses (Phase 1) must exist before notebooks can bind to them (Phase 2).

For the full execution flow with sub-steps, decision points, lift-and-shift vs. modernize paths, and error recovery, see migration-orchestrator.md.

REST API Quick Reference

All Synapse and Fabric API endpoints with request/response examples are in migration-orchestrator.md (Steps 2a–2e). Authentication tokens:

API docs: Synapse ARM · Synapse Data Plane · Fabric Items · Fabric Shortcuts · Fabric Connections · Fabric Environments

---

Migration Workload Map

Use this table to determine the correct Fabric target for each Synapse component:

Decision Tree: Which Fabric Spark Workload?

Synapse Spark workload
├── Interactive notebook with data exploration → Fabric Notebook (attached to Lakehouse)
├── Scheduled/production job → Spark Job Definition (SJD)
├── T-SQL over files/Delta → Lakehouse SQL Endpoint (no migration needed — just point to OneLake)
└── Real-time ingest → Fabric Eventstream + Lakehouse

---

T-SQL & Spark Configuration Differences

For detailed T-SQL surface area gaps (PolyBase → COPY INTO, distribution hints, result set caching) and Spark configuration mappings (pools, %%configure, runtime versions), see feature-parity.md.

Key actions: Remove DISTRIBUTION = HASH(col) hints, replace CREATE EXTERNAL TABLE with COPY INTO, replace spark.read.synapsesql() with OneLake shortcuts or JDBC. Delegate T-SQL authoring to sqldw-cli.

---

Capacity Sizing Reference

For Synapse pool → Fabric SKU mapping tables, sizing decision guide, and cost model comparison, see capacity-sizing.md.

Quick guide: Dev/test = F8–F16 with Starter Pool; standard production = F32–F64; enterprise = F128+. Use Fabric Trial (free F64, 60 days) for migration validation.

---

Must / Prefer / Avoid

MUST DO

• Preserve stored-procedure input contracts — keep every supported source input externally overridable through the Fabric Notebook Activity parameter mapped in first-cell %%configure; preserve a source default only as defaultValue, never replace a parameter use with a literal or invent a preview default, and block automatic publication when a required input has no source default
• Approve stored-procedure notebook cardinality after discovery — calculate projected workspace item demand, present 1:1, N:1, and N:N choices, and block conversion until the user provides and approves a complete mapping, target names, dependency grouping, and workspace placement; preserve per-procedure source decisions and source-block provenance under every strategy
• Audit large-procedure conversion by source block — generate deterministic per-run ledger/verifier scripts, require 100% non-overlapping source-byte coverage and a deployable disposition for every block, retry only failed blocks within the declared limit, retain audit/logging behavior by default, and publish only the exact bytes in a hash-verified ReadyForPublication package
• Use direct APIs for non-procedural phases — use SqlPackage/DMVs for discovery, Fabric REST for item management, and Fabric Livy statements for schema and Delta execution
• Replace all mssparkutils imports with notebookutils — see utility-api-mapping.md for the complete namespace table
• Replace all Linked Services with Fabric Data Connections (for external databases/services) or OneLake Shortcuts (for ADLS Gen2 / Blob storage mounts) — see connectivity-migration.md
• Replace spark.read.synapsesql() with Lakehouse shortcut reads or JDBC connections to the Fabric Warehouse SQL endpoint
• Re-test all notebooks after migration against the target Fabric Runtime version — Spark minor version differences can surface deprecated API warnings
• Externalize all workspace/item IDs — never hardcode; use pipeline parameters or Variable Libraries
• Replace pool-level library installs with Fabric Environments attached at the workspace or notebook level

PREFER

• OneLake Shortcuts over full data copies — mount existing ADLS Gen2 containers as shortcuts rather than re-ingesting data during migration
• Fabric Starter Pool for dev/test migrations — eliminates pool warm-up wait time inherent in Synapse on-demand pools
• Lakehouse SQL Endpoint as a drop-in for Serverless SQL Pool reads — point existing consumers at the endpoint with minimal query changes
• Medallion architecture for migrated data — align with Bronze/Silver/Gold patterns (see e2e-medallion-architecture skill)
• Incremental migration — migrate and validate workload by workload rather than performing a big-bang cutover
• Parameterized notebooks to allow environment promotion (dev → test → prod) without code changes

AVOID

• Do not use target notebooks as migration orchestration dependencies — generated notebooks are required outputs and are published without execution
• Do not copy-paste PolyBase CREATE EXTERNAL TABLE DDL into Fabric Warehouse — rewrite as COPY INTO or use Lakehouse for external data access
• Do not assume Synapse Linked Service connection strings are reusable — credentials and endpoints must be reconfigured as Fabric Data Connections
• Do not install libraries in notebook cells (%pip install at runtime) for production workloads — use Fabric Environments for reproducible, versioned library management
• Do not migrate Dedicated SQL Pool distribution hints (HASH, ROUND_ROBIN, REPLICATE) verbatim — remove them; Fabric Warehouse handles distribution automatically
• Do not use wasb:// or abfss://container@storageaccount.dfs.core.windows.net/ paths as primary data paths — migrate data access to OneLake abfss://workspace@onelake.dfs.fabric.microsoft.com/ paths

---

Examples

See code-patterns.md for full before/after examples. Key quick references:

mssparkutils.env → notebookutils.runtime

# Synapse
workspace = mssparkutils.env.getWorkspaceName()

# Fabric
workspace = notebookutils.runtime.context["currentWorkspaceName"]

Linked Service credential → Key Vault secret

# Synapse
conn = mssparkutils.credentials.getConnectionStringOrCreds("MyLinkedService")

# Fabric
conn = notebookutils.credentials.getSecret("https://myvault.vault.azure.net/", "my-secret")

Dedicated SQL Pool DDL → Fabric Warehouse DDL

-- Synapse (remove distribution hints)
CREATE TABLE dbo.Fact (...) WITH (DISTRIBUTION = HASH(id), CLUSTERED COLUMNSTORE INDEX);

-- Fabric Warehouse
CREATE TABLE dbo.Fact (...);

---

Feature Parity Reference

Full Synapse → Fabric feature matrix (28 features), T-SQL surface area gaps, and Spark configuration differences are in feature-parity.md.

Key gaps (⚠️/❌): spark.read.synapsesql() replaced by JDBC/shortcuts · Linked Services redesigned as Data Connections/Shortcuts · External HMS partial (migrate as shortcuts) · mssparkutils.env renamed to notebookutils.runtime · Result set caching ❌ · Workload management ❌ · PolyBase → COPY INTO

---

Migration Gotchas — Quick Reference

The full troubleshooting guide with code examples and multi-option resolutions is in migration-gotchas.md. This summary surfaces the key issues for quick scanning during migration:

---

Post-Migration: What's Next

After completing Phases 0–3 and validation, hand off to these companion skills for ongoing operations:

Agentic Exploration Workflow

Use this sequence only after a separately approved process has loaded data into Fabric Lakehouses. The Dedicated Pool to Lakehouse pattern in this skill migrates schema and code artifacts only, so it must stop after artifact validation and must not run this workflow.

For migrations that explicitly include approved data movement and data validation:

1. Discover → List schemas, tables, and row counts via Lakehouse SQL Endpoint (sqldw-cli)
2. Sample → SELECT TOP 5 on migrated tables to verify data integrity
3. Validate → Run validation checks from validation-testing.md (V1–V6)
4. Explore → Write Spark or T-SQL queries against migrated data using spark-cli or sqldw-cli
5. Build → Create Gold-layer aggregations with e2e-medallion-architecture (Bronze → Silver → Gold)
6. Consume → Build semantic models and reports with semantic-model-authoring

Companion Skill Cross-References

Variable Library for Environment Promotion

After migration, avoid hardcoded workspace/item IDs by centralizing configuration in a Variable Library item:

# Read config from Variable Library — works in notebooks
lib = notebookutils.variableLibrary.getLibrary("MigrationConfig")
lakehouse_name = lib.lakehouse_name
workspace_id = lib.workspace_id

# ❌ WRONG — .get() does not exist
# notebookutils.variableLibrary.get("MigrationConfig", "lakehouse_name")

• Use Value Sets (valueSets/dev.json, valueSets/prod.json) to promote across environments without code changes
• Boolean values are returned as strings — compare with .lower() == "true", not bool()
• In Data Pipelines, reference via @pipeline().libraryVariables.<name> (not @variables())
• Full Variable Library patterns → see common/notebook-authoring/context-and-params.md § Variable Library