Limits & performance
Set Fabric capacity, concurrency, and workload guardrails for reliable operations.
Operational boundaries
Compute sizing
Fabric Warehouse compute is tied to the workspace's assigned Fabric capacity (an F-SKU or trial capacity), not a per-query warehouse you resize yourself the way you would on Snowflake. Size the assigned capacity for your workload's concurrency and data volume at the Fabric-admin level; Vulcan has no lever to resize Fabric capacity from the project config.
Concurrency
Start with concurrent_tasks: 2 for daily incrementals. Because there are no explicit transactions and the overwrite strategy is DELETE_INSERT rather than MERGE for time-range/partition kinds, avoid overlapping runs against the same table. A second run's DELETE can race a first run's INSERT with no transactional isolation to fall back on. Use concurrencyPolicy: Forbid for every scheduled workflow on this engine.
Scheduling and latency floor
Schedule after upstream lands, set timezone: UTC, set endOn 1 to 2 years out, and use concurrencyPolicy: Forbid. The latency floor includes: warehouse SQL endpoint connection setup (Entra token acquisition plus ODBC connect), no index-assisted point lookups (every filtered query is closer to a scan than it would be on an indexed MSSQL table), and asynchronous warehouse creation/drop latency (seconds to tens of seconds) if your workflow provisions warehouses dynamically.
Performance reference
Because there is no user-tunable indexing, sizing conversations for Fabric center on query shape and capacity tier rather than index design. Track plan/run duration per model, semantic query wall time, and rows touched per incremental window. When concurrency or latency degrades, fix in order: remove cross-join-shaped queries, pre-aggregate high-cardinality COUNT(DISTINCT ...) measures into a mart, stagger schedules so DELETE_INSERT-based incrementals never overlap on the same table, then escalate the Fabric capacity tier as a platform-level change. Reproduce measurements against your own workspace capacity before quoting them externally. Fabric capacity tiers behave differently from a dedicated SQL Server instance under load.
Next step
Use Troubleshooting to diagnose connection, deployment, and runtime failures.
Last updated
Was this helpful?