Capacity management¶

CSA Loom doesn't have a single F-SKU billing unit (Fabric does). Instead, the Loom Console synthesizes a CU-equivalent dashboard from the underlying-service consumption.

The CU-equivalent model¶

Underlying Azure consumption	CU equivalent
Databricks DBU	1 DBU ≈ 16 CU
Synapse Serverless DPU	1 DPU ≈ 8 CU
ADX vCore-second	1 vCore-second ≈ 1/60 CU
Power BI Premium memory-MB-hour	1 MB-hour ≈ 1/1024 CU
AOAI TPM (tokens per minute)	50,000 TPM ≈ 1 CU
Function invocations	negligible (~0 CU)
ADLS Gen2 storage	not counted (storage-only, not compute)

These coefficients are approximations calibrated against Fabric F-SKU F64 / F128 published throughput. They give customers a single number to plan against, even though the underlying billing remains separate.

Capacity scaling¶

Console "Admin → Capacity" pane lets the customer adjust:

Service	Scaling axis
Databricks SQL Warehouse	Small / Medium / Large / X-Large (Serverless tier auto-scales within bound)
Databricks classic clusters	Min/max nodes per cluster policy
ADX cluster SKU	Dev / Standard_D11_v2 / Standard_E16ds_v5 / etc.
Power BI Premium F-SKU	F4 / F8 / F32 / F64 / F128 / F512
AOAI deployment TPM	Per-deployment quota allocation
APIM tier	Premium throughput tier

Scaling actions: - Up: instant (no downtime — Azure-native scaling) - Down: scheduled (avoids interrupting in-flight workloads) - Audit: every scale action logged → Activity Log → Sentinel (Gov)

Pause / resume patterns¶

Most expensive consumption is Spark + ADX:

Service	Pause method	Cost savings
Databricks all-purpose clusters	Auto-pause after idle minutes (cluster config)	~70% DBU savings overnight
Databricks SQL Warehouse	Auto-suspend after idle minutes	~70% DBU savings
ADX cluster	Stop cluster (CLI / Azure portal)	Compute cost → 0; storage cost continues
Power BI Premium capacity	Pause-capacity (Azure portal)	Full F-SKU $ → 0 while paused
AOAI provisioned throughput	De-allocate provisioned units	Saves vs PAYG when not in use

Console Admin pane has one-click pause for the entire DLZ (pauses Databricks + ADX + Power BI together; reads remain free; writes fail until resume).

Capacity overrun monitoring¶

Console "Monitoring → Capacity" pane shows:

CU-equivalent over last 24h (sum of DBU + DPU + vCore-seconds + Power BI memory + AOAI TPM, weighted)
Per-service breakdown (Databricks / ADX / Power BI / AOAI)
Forecast vs budget
Throttling events (per-service 429 / capacity-throttled events)

Alert rules deployed by platform/fiab/bicep/modules/admin-plane/monitoring.bicep: - CU-equivalent > 80% of budget for 1h → low-sev alert - CU-equivalent > 100% of budget for 1h → high-sev alert - Per-service 429 rate > 5% → service-specific alert

Forecasting¶

fiab-capacity-forecast (script in platform/fiab/bicep/scripts/) projects CU-equivalent usage from the last 30 days. Run before adjusting reservation purchases.

Reservations¶

Service	Reservation pattern
Databricks DBU	DBU Commit Units (1-year / 3-year); ~30-40% savings
ADX	Reserved instances (1-year / 3-year); ~30% savings
Power BI Premium F-SKU	Annual commit; ~40% savings vs PAYG
AOAI provisioned throughput	Provisioned managed throughput (PTU) reservations
Synapse Serverless	No reservations — pure pay-per-query

Per-boundary considerations¶

Boundary	Gov premium delta vs Commercial
GCC	Same Azure Commercial pricing
GCC-High / IL4	~10-25% above Commercial list
IL5 (v1.1)	Same as GCC-H + HSM-CMK overhead

Runbook¶

Capacity overrun