Azure Data Factory with a self-hosted integration runtime: production setup checklist

#Host

• Dedicated Windows Server VM (not a domain controller, not the database server), 8+ cores, 16–32 GB RAM for typical loads.
• Close to the data sources (same data centre/VNet) to keep latency low.
• Java runtime installed if you read/write Parquet or ORC on the SHIR (64-bit JRE/OpenJDK).
• Antivirus exclusions for the SHIR installation and temp folders (per Microsoft guidance) if scans slow copies.

#Install and register

# after downloading IntegrationRuntime_<version>.msi
msiexec /i IntegrationRuntime.msi /quiet
& "C:\Program Files\Microsoft Integration Runtime\5.0\Shared\dmgcmd.exe" -RegisterNewNode "<authentication-key>"

The key comes from ADF Studio → Manage → Integration runtimes → New → Self-Hosted.

#Network

• Outbound HTTPS 443 to Azure (*.servicebus.windows.net, *.frontend.clouddatahub.net, the factory's region endpoints, and storage endpoints you copy to). No inbound ports needed.
• Proxy configured in the SHIR configuration manager if outbound goes through a proxy.
• If the factory uses a managed VNet/private endpoints, the SHIR still connects outbound; check private DNS for storage endpoints from the SHIR host.

#Credentials

• Linked service secrets in Azure Key Vault (ADF managed identity with Get/List on secrets).
• Windows auth to SQL Server via a service account; the SHIR service runs as that account only if required (default is NT SERVICE\DIAHostService).

#Performance

• Concurrent jobs limit on the node (default based on cores/memory) tuned after observing CPU/memory.
• Copy activity: parallel copies and staged copy for large loads.

#Updates

• Auto-update enabled (Microsoft supports only recent SHIR versions) with a time window, or a scheduled manual update process.

#Monitoring

• ADF Monitor → Integration runtimes: status, CPU, memory, concurrent jobs.
• Azure Monitor alert on IR status not Online and on pipeline failures.
• SHIR event logs (Integration Runtime log in Event Viewer) collected by the monitoring agent.

#Single-node limits

For production pipelines with SLAs, add a second node (see the SHIR HA post). Up to four nodes can share one logical SHIR.