Reducing Azure costs: sizing, networking and storage
Lazy Read
Before reducing resources or buying commitments in Azure, compare billing with actual usage and service requirements. This article explains what to review across machines, plans, networking and storage, and how to propose measurable changes with a rollback path.
- Distinguish actual costs, amortization and committed capacity.
- Check utilization and peaks before changing resources.
- Validate savings against performance, security and recovery.
Related links
A VM enlarged for a launch, several App Service plans ready for growth and test data still stored months later. Each decision may make sense on its own. Over time, the bill can stop matching what the product needs.
Optimizing Azure costs requires connecting spending, workloads and service requirements. Useful savings preserve availability, performance and security; their size depends on the starting point. This walkthrough identifies what to review before reducing capacity or committing to spending.
Understand what each service is paying for
Break down spending by subscription, resource group, service and region in Microsoft Cost Management. Tags help connect it to teams and environments, but a resource without an owner needs investigation, not automatic deletion.
Distinguish actual and amortized costs. The former reflects charges; the latter spreads commitment costs across their term and helps examine utilization. Treating both views as equivalent can obscure unused commitments.
Complete the picture with inventory and Azure Advisor recommendations. These are signals to check against the service, not decisions to apply without context.
Size resources with business peaks in view
Review CPU, instrumented memory, disks, networking and latency. Azure Monitor distinguishes platform metrics from guest-system data; the latter may require an existing agent. A read-only audit should not install one.
A low average may hide a critical export that saturates a VM for a few minutes. Before resizing, check when it happens, which resource limits it and how much headroom the product requires.
If the task can be separated, compare Container Apps Jobs, Batch or a temporary worker. The goal is to avoid letting an occasional task determine permanent capacity. Include startup, retries, storage and operations in the comparison.
App Service, AKS and the components we keep running
In dedicated compute tiers of App Service, applications in the same plan share its instances. Review size, instance count and combined usage before retaining several lightly used plans or consolidating them. Sharing capacity also means sharing limits and may change isolation.
In AKS, examine nodes, resource requests, disks, load balancers and observability. Orchestration may be justified by deployment or scaling requirements; it also introduces operational work that belongs in the cost calculation.
Separate services when they need to evolve or scale differently. If they always share the same lifecycle, investigate whether the added traffic, gateways and dependencies deliver measurable value.
Networking: review costs while preserving controls
NAT Gateway charges for resource hours and data processed; outbound transfer may add charges. Private Link also has pricing conditions to include. Review traffic paths, associations and the purpose of each component.
Include Application Gateway, firewalls, VPNs and interregional transfers where they are part of the architecture. A quiet gateway may provide backup connectivity, and apparently duplicated networking may separate trust boundaries.
Before simplifying, demonstrate that routes, connectivity, security and resilience remain intact. A component having a cost does not establish that it is unnecessary.
Storage, logs and work we could avoid
In Blob Storage, compare access frequency, retrieval and retention before changing tier. Cool and Cold remain online; Archive requires rehydration and can involve hours of waiting. Early deletion or tier changes can incur charges.
Lifecycle policies automate transitions and expiration, but require a clear retention and recovery policy. Also review versions, snapshots, disks without associations and backups, preserving the copies needed to restore the service.
In Log Analytics and Application Insights, connect ingestion, sampling and retention to operational questions. Avoid losing important security or error signals when reducing volume.
Front Door caching can avoid origin requests for suitable content. Measure hits and work avoided; test expiration, invalidation and separation of public and private content. Caching one user's response for others would turn an optimization into an incident.
Commit after understanding demand
Reservations and savings plans offer discounts under different conditions. An attractive discount does not compensate for paying for capacity you no longer need.
Estimate steady demand after adjusting sizes and architecture. Review utilization, coverage, duration, flexibility and benefit scope. Calculate proposed costs with existing commitments and avoid adding incompatible savings over the same usage.
Turn analysis into a measurable change
Compare several representative weeks of billing and utilization, extending the window for seasonal workloads. For each proposal, document the resource and owner, evidence, missing data, comparable net cost, dependencies and risks.
Agree on error, latency and availability measurements before and after the change, and on a rollback path. The audit can use read-only access; applying its recommendations is a separate operation requiring its own authorization. No inventory guarantees a migration without interruption.
The Azure audit skill 0.1.0 organizes this process with an agent and includes a bounded metadata, cost and CPU collector. It generates private local reports with explicit coverage gaps. Its cost query uses POST as a read operation; it does not modify infrastructure. Automated inventory still needs business context, sufficient metrics and validation of proposed savings.