hdinsight-migration

Telemetry — MANDATORY. Every api.fabric.microsoft.com call must carry x-ms-fabric-skill: hdinsight-migration (az rest: --headers "x-ms-fabric-skill=hdinsight-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. HDInsight has no mssparkutils or dbutils equivalent — notebookutils is net-new capability being introduced
4. HiveContext and SQLContext are legacy Spark 1.x/2.x APIs — Fabric uses Spark 3.x SparkSession exclusively
5. wasb:// paths are deprecated and require a Storage Account key or SAS — replace with OneLake shortcuts
6. Fabric cannot read or shortcut hdfs:// directly. Its Pipeline HDFS connector supports Anonymous authentication only; export or bridge Kerberos sources to supported storage first.

HDInsight → Microsoft Fabric Migration

Prerequisite Knowledge

Read these companion documents before executing migration tasks:

• 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, Spark job execution

For notebook and Lakehouse creation, see spark-cli. For Fabric Warehouse DDL/DML authoring, see sqldw-cli.

Table of Contents

Migration Workload Map

SparkSession & Context API Changes

HDInsight Spark clusters often use legacy Spark 1.x / 2.x API styles. Replace all of these with the unified SparkSession:

In Fabric notebooks, spark (SparkSession) and sc (SparkContext) are pre-instantiated — do not call SparkContext() or SparkSession.builder...getOrCreate() at the top of migrated notebooks.

Oozie → Fabric Pipelines

Map Oozie workflow actions to Fabric Data Pipeline activities:

Delegate to spark-cli for notebook and SJD creation after mapping pipeline activities.

Introducing notebookutils

HDInsight Spark had no built-in utility framework equivalent to mssparkutils or dbutils. When migrating to Fabric, introduce notebookutils for common operations:

Spark Configuration Differences

Must / Prefer / Avoid

MUST DO

• Replace all wasb:// / wasbs:// paths with OneLake abfss:// paths or OneLake Shortcuts — wasb:// requires storage account keys which are not the Fabric-preferred auth model
• Replace HiveContext, SQLContext, and standalone SparkContext() — use the pre-instantiated spark session in Fabric notebooks
• Migrate Hive DDL (STORED AS ORC, LOCATION, TBLPROPERTIES) to Delta Lake DDL — see hive-to-delta.md
• Introduce notebookutils for file system operations, secret retrieval, and child notebook orchestration where HDInsight used custom scripts or direct API calls
• Replace Oozie XML workflows with Fabric Data Pipelines — see § Oozie → Fabric Pipelines
• Align library management to Fabric Environments — remove bootstrap.sh, conda envs, and runtime %pip install patterns for production workloads
• Treat hdfs:// as source-cluster-only — export or ingest into ADLS Gen2 or OneLake; do not claim that a gateway adds Kerberos support to Fabric's Anonymous-only HDFS connector

PREFER

• OneLake Shortcuts over copying data — mount existing ADLS Gen2 containers as shortcuts to avoid re-ingestion during migration
• Delta Lake for all tables migrated from Hive ORC/Parquet — ACID guarantees, time travel, and schema enforcement improve data quality
• Fabric Starter Pool for initial migration validation — no pool configuration overhead, fast session startup
• Lakehouse schemas (database namespaces) for organizing migrated Hive databases — one schema per Hive database within a single Lakehouse
• Medallion architecture for restructuring migrated data layers during migration — align Bronze/Silver/Gold with raw Hive → validated Delta → serving Gold patterns

AVOID

• Do not use SparkContext() or HiveContext() constructors in Fabric notebooks — they conflict with the pre-instantiated spark session and will raise errors
• Do not use hive-site.xml or external Hive metastore configuration — Fabric's Delta Lake-backed Lakehouse IS the metastore
• Do not assume YARN queue mappings translate to Fabric pools — re-design resource allocation based on Fabric Spark pool SLAs
• Do not attempt to run Oozie shell actions or Java MapReduce jobs directly in Fabric — these must be refactored (see § Oozie → Fabric Pipelines)
• Do not use %sh magic for file system operations in production notebooks — use notebookutils.fs.* for portability and OneLake token-based auth

Examples

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

Legacy context → Fabric pre-instantiated session

# HDInsight (remove entirely)
from pyspark.sql import HiveContext
hc = HiveContext(sc)

# Fabric — use pre-instantiated spark directly
df = spark.sql("SELECT * FROM sales.fact_orders")

WASB path → OneLake path (after shortcut creation)

# HDInsight
df = spark.read.parquet("wasb://raw@myaccount.blob.core.windows.net/orders/")

# Fabric
df = spark.read.parquet("Files/raw/orders/")

Hive DDL → Delta DDL

-- HDInsight
CREATE TABLE sales_db.fact_orders (...) STORED AS ORC LOCATION 'wasb://...';

-- Fabric
CREATE SCHEMA IF NOT EXISTS sales_db;
CREATE TABLE sales_db.fact_orders (...) USING DELTA;