Frequently Asked Questions

General, architecture, and logistics questions for the AgentOps Control Tower RVAS. Coaches should also review the per-challenge guides in coach/.

Logistics#

How long is the RVAS? Plan for 1.5–2 days. Challenges 0–5 are the core path; Challenge 6 is a stretch for teams that move fast. Each challenge is scoped to roughly 1.5–2.5 hours including discussion.

Do we build everything from scratch? No. The resources/ directory is a complete reference implementation — Bicep, application code, Fabric setup scripts, PySpark notebooks, pipelines, and a semantic model. The intent is that you deploy the Azure foundation (Challenges 1–2) quickly and spend your real energy on the Fabric Control Tower (Challenges 3–6). Coaches decide how much of the reference to reveal based on your team's level.

Can teams share one Fabric capacity? Yes. One trial capacity (or paid F-SKU) can back many workspaces. The capacity admin assigns each team's workspace to the shared capacity. Give each team its own workspace for isolation.

What if our network blocks egress? Flag it before Day 1. You need outbound HTTPS to Azure, *.fabric.microsoft.com, and package registries (PyPI, npm, MCR/Docker Hub). Azure Cloud Shell is a good fallback for CLI work.

Environment & access#

Do we need multiple Azure subscriptions? One is enough. Multiple is fully supported and realistic — e.g., workload in one subscription, cost exports scoped at the billing account. Record your subscription IDs in Challenge 0.

Which region should we use? One that offers both your chosen model (gpt-4o / gpt-4o-mini) and Microsoft Fabric. East US 2, West US 3, and Sweden Central are safe defaults. Keep all resources in one region to avoid egress charges with OneLake shortcuts.

What happens when the Fabric trial expires? After 60 days, trial-capacity access is revoked and assigned workspaces are reassigned. Content stays in OneLake for 7 days and can be reactivated by assigning the workspace to a paid F or Power BI Premium capacity. For a 2-day event this is a non-issue — just don't let a long gap pass before revisiting.

Do we need Power BI Desktop? No — you can author the semantic model and reports in the Fabric/Power BI service. Desktop (Windows) is convenient for offline model edits but optional.

Architecture & design#

Why a medallion (Bronze/Silver/Gold) architecture instead of one transformation? Separation of concerns. Bronze preserves raw data for reprocessing and audit. Silver standardizes across sources so consumers get a consistent interface. Gold pre-computes aggregations so dashboards are fast. If a transformation bug appears, reprocess from Bronze without re-ingesting from source.

Why Fabric instead of Synapse or Databricks? Fabric unifies data engineering, warehousing, and BI on a shared OneLake. For this scenario the tight integration between Lakehouse, notebooks, pipelines, Direct Lake, and Data Activator removes data movement and simplifies the architecture. Synapse and Databricks are great alternatives for the medallion layer — the pattern is portable.

Why OneLake shortcuts and Mirroring instead of copying data in? Shortcuts let Fabric read ADLS Gen2 telemetry in place — zero duplication, zero intra-region egress. Mirroring replicates Cosmos DB conversations into OneLake Delta tables near-real-time with no ETL to build or operate. Together they minimize cost, latency, and maintenance.

What makes this a "Control Tower" and not just a dashboard? It correlates signals that normally live in silos — cost (Cost Management), reliability (Azure Monitor), and agent behavior (App Insights + Cosmos DB) — into shared Gold data products, and then acts on them with real-time alerts. One pane of glass plus automated response.

How does cost attribution / chargeback work? The Silver/Gold layers join FOCUS cost data with resource tags and the agent namespace (<org>.<domain>.<agent>.<version>). That lets you attribute blended cost (model tokens + Cosmos RUs

• compute + APIM) to a team, use case, or client. Challenge 6 builds this out with RLS.

Performance & scale#

How much data can this handle? It scales with Fabric capacity. F2/F4 handles tens of millions of rows comfortably for the RVAS. For enterprise volumes (billions of cost records), use F64+, partition Delta tables by month, and Z-order on frequently filtered columns (subscription, date).

What's the end-to-end latency? ADLS Gen2 data follows the pipeline schedule (the reference daily pipeline runs ~06:00 UTC). Cosmos DB mirrored data is typically available within a minute of the write. Direct Lake reflects the latest Delta state after each pipeline run — no scheduled dataset refresh needed.

How long does the end-to-end pipeline take? For a typical 3-month window (a few million cost rows, hundreds of thousands of metric points, tens of thousands of logs), the full Bronze→Gold pipeline runs in ~10–15 min on F2; the daily incremental finishes in 2–3 min.

Security & cost#

How is authentication handled? Managed identities and Entra ID throughout — the agent services use managed identity to reach Azure OpenAI and Cosmos DB; Fabric uses workspace identity / managed identity to reach ADLS Gen2; Mirroring authenticates via Entra ID. Optional CI/CD uses GitHub OIDC federated credentials (no stored secrets).

Is the data encrypted? Yes. ADLS Gen2 and OneLake encrypt at rest (Microsoft-managed keys by default; customer-managed keys available). Cosmos DB is encrypted at rest. All traffic uses TLS 1.2+.

What will this cost us? Low for an RVAS: Fabric is free on trial; the agent workload runs a few dollars/day with gpt-4o-mini, serverless Cosmos DB, and scale-to-low Container Apps (APIM Developer tier is the main fixed cost); the landing zone is inexpensive. Tear everything down with azd down --purge and end the Fabric trial / delete the workspace when finished.

Can we pause Fabric to save money? Yes (paid capacities). Pause outside working hours; schedule pipelines within the active window. OneLake data persists while paused. Trial capacity needs no pausing — it's free.

Troubleshooting quick hits#

More detail lives in each challenge's coach guide.