Custom Vision

Security for Custom Vision starts with knowing who can view, change, export, or act on the evidence. Use least-privilege Azure RBAC, Microsoft Entra identities, managed identities where relevant, private or restricted data paths, and logged approval workflows. Avoid exposing training images, prediction images, endpoint keys, project names, tags, model performance reports, and manufacturing or customer identifiers in dashboards, tickets, exports, repositories, or scripts. For Custom Vision, images can reveal products, locations, people, or regulated records, so storage, labeling, endpoint access, and logs need tight control. A secure design records owner, scope, allowed readers, change authority, retention expectations, break-glass path, and review cadence so troubleshooting does not become a reason for broad access or unmanaged data sharing.

Cost for Custom Vision shows up through training and prediction transactions, image storage, labeling effort, edge exports, duplicate projects, and migration work for future replacement. Measure the signal before changing the setting or blaming the platform, and track ownership, exceptions, and review dates. A cheap configuration for one workload can be expensive for another when traffic patterns, retention, tagging, query shape, or ownership boundaries change. Use tags, budgets, alerts, exports, and per-scope dashboards so product owners can see which behavior drives spend. The strongest cost review connects dollars to a real behavior, such as requests, storage, idle capacity, alerts, shared services, or untagged resources.

Reliability for Custom Vision depends on predictable behavior during spikes, month-end processes, deployment changes, regional events, or dependency failures. Test training iteration status, prediction endpoint health, quota availability, model drift, batch retry handling, and documented migration milestones with production-shaped data, realistic time windows, and documented recovery steps. Operators should know which symptoms indicate stale data, missing tags, throttling, bad filters, alert noise, or resource pressure. Include rollback or mitigation steps before changing production resources or cost controls, because the setting often affects more than one team. Review the runbook during planned tests. The goal is not only availability; users need correct signals, acceptable response time, and a known path when conditions change.

Performance for Custom Vision is measured through prediction latency, image size, endpoint throughput, precision and recall, class balance, model version, and edge or cloud execution time. Review the signal with production-shaped data instead of tiny development samples or one-day cost snapshots. Azure Monitor metrics, Cost Management views, CLI output, SDK diagnostics, and portal evidence should tell the same story. Tune the design only after separating application delays, billing latency, tagging gaps, and configuration drift. A good performance fix reduces latency, noise, or operator effort without weakening security, correctness, allocation accuracy, or recovery. Capture baseline, change, and rollback evidence together. Re-test after deployments because traffic, tags, indexes, and usage patterns can shift the result.

Operations for Custom Vision should be repeatable enough that a second engineer can verify the same facts without tribal knowledge. Keep project owners, labeled datasets, iteration history, published endpoints, quota checks, prediction monitoring, and migration backlog documented with deployment source, owner, change history, dashboard links, and escalation contacts. Use read-only Azure CLI checks, portal review, Azure Monitor or Cost Management views, and export evidence to compare intended state with live behavior. Runbooks should say what is safe to inspect, what requires approval, and what evidence must be captured before and after a change. Review the record after each production change. Good operations make the term a checked production control, not a hidden implementation choice.