Weekly Azure Roundup: Sovereign AI, Metrics, and Migration Planning

by TechHub

Welcome to this week's Azure roundup, covering sovereign AI infrastructure, new monitoring capabilities, and upgrade planning. Azure Local announcements expand options for local inference and disconnected operations, while generally available registry and Azure Monitor metrics help teams investigate traffic, throttling, storage usage, and telemetry growth. We review the December retirement of Foundry workflows and the September 2027 Java support deadline for App Service and Functions, then explore agent development, database modernization, and private connectivity, distinguishing available features from previews and planned releases.

This Week's Overview

Java support deadlines require upgrade planning

Azure App Service and Azure Functions will end support for Java 8, 11, and 17 on September 1, 2027. Existing applications will continue to run, but those Java versions will stop receiving security updates, improvements, and related customer support. Teams should treat the deadline as a support and security boundary, not an application shutdown date.

The Microsoft Build of OpenJDK lifecycle is a separate consideration: September 2027 is the earliest listed end-of-support date for Java 11 and 17, and those targets may be extended. Microsoft supports Java 8 on Azure through Eclipse Temurin rather than a Microsoft Build of OpenJDK 8. Modernize CLI can assess applications, help plan upgrades, and apply changes for review.

Azure Container Registry makes traffic and throttling visible

Azure Container Registry's new metrics give operators a shared view of outbound traffic and registry activity. The accompanying engineering articles explain how the service measures downloads that leave the registry's direct serving path.

Generally available metrics and dashboards

Data Transfer and Transactions metrics are generally available automatically for every registry, with no additional configuration or charge for the metrics themselves. Data Transfer measures outbound bytes by API name and region, while Transactions counts API operations and adds a response-type dimension. Filtering Transactions to ResponseType=Throttled identifies which APIs and regions are affected.

The integrated Grafana dashboard brings together transaction, outbound-transfer, throttling, and storage panels without requiring a separately managed Grafana instance. Operators can use these views to distinguish increased traffic from throttling rather than treating every registry slowdown as the same problem.

How download accounting works

Data Transfer attributes egress to GetManifest, GetBlob, GeoReplication, and ImportSource, including service-initiated transfers, with a PT1M time grain. GetBlob downloads typically redirect clients to Azure Storage, so the registry serving path cannot directly observe the transferred bytes. The Storage Activity Platform reduces telemetry arrival times from days to minutes in many regions without affecting that serving path.

The implementation assigns each registry's activity to one internal metrics shard across regions and components. Failed requests count as Transactions but not Data Transfer, and redirects and downloads are counted separately while delivered bytes are counted once. These distinctions matter when comparing request counts with bandwidth, particularly because a one-minute metric grain does not mean every underlying event arrives within one minute.

Sovereign infrastructure expands local AI and customer control

The sovereign-cloud announcements cover local inference, disconnected operations, hardware isolation, and control over encryption keys. Availability varies across these capabilities, so deployment planning needs to distinguish generally available components from previews and Ignite targets.

Local inference and disconnected infrastructure

Foundry Local runs supported models on Azure Local through Azure Arc-enabled Kubernetes, with CPU- and GPU-backed deployments and OpenAI-compatible APIs. Azure Local disconnected operations uses a local control plane to run supported infrastructure and services without continuous public-cloud connectivity. Public-preview rehoming adds a phased path from Azure-connected infrastructure to disconnected environments, with user approvals and secure backups.

The articles describe feature readiness differently: the dedicated Foundry Local article calls the model catalog and Agentic Retrieval generally available, while the broader Sovereign Private Cloud article describes agentic retrieval, tool calling, and an MCP server as public preview. The dedicated article places agents, enterprise connectors using Model Context Protocol (MCP), and configurable AI guardrails in preview. Teams should verify the status of the particular retrieval and agent capabilities they intend to deploy rather than assuming a single availability label covers the whole stack.

The private-cloud roadmap also places DocumentDB and Blob Storage on Azure Local in public preview at Ignite. SQL Server on Azure Local general availability was announced at FabCon.

Hardware isolation and attestation

Confidential virtual machines on Azure Local are generally available, using hardware-encrypted memory and attestation to protect data from host environments and infrastructure administrators. The GTC Berlin announcement separately says Foundry Local on Azure Local is expected to become generally available at Microsoft Ignite. That platform milestone should not be confused with the availability of individual model, retrieval, or agent features.

The Intel-focused architecture guidance describes Azure Local on Intel Xeon 6 and supported confidential VM configurations that use Intel TDX to isolate workloads while in use. Hardware-backed attestation supplies evidence that an environment meets workload security requirements. Intel AMX supports CPU-based AI acceleration, while Intel OpenVINO supports model optimization and deployment across supported infrastructure.

Development workflows on Azure Local

GitHub Enterprise Local keeps repositories and core development workflows on customer-controlled Azure Local infrastructure. Workload Orchestration guides deployment of the core infrastructure and GitHub Enterprise Server appliance, with deployment secrets stored in Azure Key Vault.

Local hosting does not remove every external dependency or operational responsibility. Connected GitHub Copilot use requires a GitHub.com identity and a GitHub Copilot license, while administrators using Foundry Local remain responsible for approved models, security, certificates, and access controls.

Scheduled recovery points for disconnected operations

Azure Local 2609 introduces scheduled backups for the disconnected operations control plane, with twice-daily, daily, weekly, and monthly frequencies. Operators use PowerShell to manage default or named configurations, inspect the next planned run, and pause, resume, update, or remove schedules. On-demand backups remain available for additional recovery points.

Public-cloud key ownership and access oversight

External Key Management with Azure Managed HSM is generally available, allowing customers to retain root encryption keys in customer-operated hardware. Data Guardian and Production Touch Ledger add oversight and tamper-evident records for qualifying Microsoft engineer access in EU and EFTA regions. The updated Sovereign Landing Zone implementation supports Bicep, Terraform, and Azure Verified Modules.

Sovereign Control Panel remains in private preview, with public preview planned for Microsoft Ignite and delivery timelines subject to change. Azure Confidential Clean Rooms is in limited preview for approved Spark SQL analysis inside hardware-protected environments. These preview services have different readiness levels from the generally available key-management capability.

Protecting sensitive inference beyond the GPU

Azure NCCads H100 v5 confidential VMs combine fourth-generation AMD EPYC processors and NVIDIA H100 Tensor Core GPUs, extending the trusted execution environment across AMD SEV-SNP and the GPU's confidential mode. Microsoft Azure Attestation assesses VM state, and Azure Key Vault Managed HSM supports policy-controlled key release after successful attestation.

The Protopia architecture transforms prompts and context through a Stained Glass Proxy inside the data owner's trust boundary before sending protected representations to Azure. Its central caution is that GPU attestation alone does not verify how gateways, caches, logs, and traces handle sensitive requests. Security reviews therefore need to cover the full inference path, not only the compute boundary.

AI-assisted operations adds infrastructure context

Azure Copilot for Azure Local and Azure SRE Agent address related operational work, but at different stages of availability. The former is an announced troubleshooting experience, while the latter adds integrations and guidance for existing investigation workflows.

Azure Copilot for Azure Local targets early 2027 preview

Azure Copilot for Azure Local translates natural-language questions into guided infrastructure troubleshooting workflows. Supported scenarios include health, connectivity, network configuration, failed deployments, update readiness, and diagnostic collection.

Investigations gather context, run diagnostics, analyze findings, and recommend next steps while operators remain in control. Public preview is planned for early 2027, so this announcement describes an upcoming capability rather than a feature operators can adopt now.

SRE Agent connects investigations with developer tools

Azure SRE Agent Canvas brings operational findings and context into GitHub Copilot Apps and Copilot CLI. The October update also supports private Azure DevOps repositories as a marketplace for plugins and skills, and supplies ServiceNow guidance for incident-type response plans that trigger investigations within existing processes.

Two private Splunk MCP samples cover same-region and cross-region connections without opening Splunk to inbound internet traffic. Updated security guidance addresses execution isolation, secretless credentials, customer separation, encryption, network controls, logging, and data residency. These details help teams evaluate how investigation tools fit their access and network boundaries.

Agent development: workflow retirement, browser tools, and deployment paths

Agent teams face an immediate orchestration deadline alongside new browser automation and development options. The practical distinction is between moving existing workflow execution, adding controlled tools, and choosing where developers build and test agents.

Move Foundry workflow orchestration before December

Microsoft Foundry workflows retire on December 1, 2026, ending support for the visual workflow designer and for running workflows in the Foundry portal. Azure Logic Apps provides an orchestration alternative, triggering Foundry agents through API calls, schedules, or events from more than 1,400 connectors. It supplies branching, retries, timeouts, parallel execution, and run history around agent calls.

Prompt Agents use configured instructions, models, and tools, while Hosted Agents run custom containerized code with Foundry-managed hosting and scaling. Selecting an existing Hosted Agent from a Foundry project for workflow invocation is planned, not currently supported. Migration designs should account for that limitation rather than assuming every agent invocation path is already available.

Browser automation becomes generally available

Microsoft Foundry's Browser Automation Tool is generally available and uses Playwright Workspaces to provision isolated cloud browser sessions. Agents can reach public websites and private applications behind an Azure virtual network, provided private connectivity is configured during workspace provisioning. Foundry toolboxes expose MCP-compatible endpoints for sharing browser tools across Prompt agents and Hosted agents.

Live View lets people observe browser activity, and Take control supports temporary human interaction before handing the session back to the agent. Controls include sensitive-action approvals, tool allow lists, managed identities, and Azure role-based access control. Sensitive trace protection remains a preview capability, an important distinction when evaluating what browser sessions may record.

Build grounded agents in VS Code or the Copilot app

Julia Muiruri's Visual Studio Code walkthrough uses GitHub Copilot and the Foundry Toolkit extension to build a game-recommendation agent grounded in a real product catalog. It covers model selection, testing, deployment to Azure AI Foundry, and connection to a website through a secure proxy.

A separate GitHub Copilot app walkthrough starts from a GitHub issue prompt to build a Tailspin Toys backer concierge agent. Foundry Skill and Foundry Canvas support resource provisioning, model deployment, and hosted-agent development, while an Azure Functions proxy protects credentials used by the website chat widget. The workflow includes testing and grounding checks, with a linked self-paced workshop for hands-on practice.

Use a controlled model endpoint from the Copilot SDK

The .NET 10 bring-your-own-model sample retains the GitHub Copilot SDK runtime while routing inference to a user-controlled OpenAI-compatible endpoint through ProviderConfig and the Responses API. Without an API key, AzureCliCredential reuses Azure CLI authentication and requires the Cognitive Services OpenAI User role.

Authentication callbacks for this approach are experimental in the Copilot SDK. The sample disables tools, skills, configuration discovery, and Git operations to isolate the model request, so it demonstrates endpoint integration rather than a complete tool-enabled agent.

Observability adds usage, capacity, and allocation detail

The monitoring updates make it easier to inspect storage growth, telemetry ingestion, and GPU memory pressure. They also have different pricing and preview conditions, which matter when enabling them across an estate.

Container-level storage metrics reach general availability

Azure Monitor advanced platform metrics are generally available, with Azure Storage as the first supported resource. Container Blob Capacity and Container Blob Count expose storage usage and object counts at container level. In Metrics Explorer, users can select Apply splitting and Container name, and use the metrics for alerts.

This is a paid capability enabled per storage account through the Azure portal, Azure CLI, or PowerShell. Availability includes Fairfax, Mooncake, and all Azure public regions supported during public preview. Teams should distinguish its billing model from the no-additional-cost registry metrics announced separately.

Workspace insights explain ingestion changes

Metrics usage insights is generally available wherever Azure Monitor workspace is generally available in public regions. The Ingestion Volume Change dashboard compares two points in time and identifies which metrics contributed to the change. Built-in views show active time series and event ingestion rates against workspace limits, and identify metrics that have not been queried recently.

Teams can enable the feature during workspace creation or through diagnostic settings on an existing workspace. This gives operators a way to investigate ingestion growth and review unused metrics before changing collection settings.

Storage Discovery separates active data from retained capacity

Advanced capacity insights is in preview and included in the Azure Storage Discovery Standard plan at no additional cost. After enabling Capacity details in a Discovery workspace, users can examine storage tier alongside last access, creation, or modification time to identify inactive data.

Storage composition separates live data from previous versions, snapshots, and soft-deleted objects. Retention charts expose potential early deletion charges and recoverable deleted data that still consumes capacity, helping teams avoid equating logical deletion with immediate storage savings.

Private endpoint compression reduces address consumption

The Data Collection Endpoint (DCE) IP Endpoint Compression preview lets all DCEs in a region share three regional private IP addresses. The example reduces 20 DCEs from 60 private IPs to three, approximately 95% fewer. Compression applies automatically to newly created DCEs in preview-enabled subscriptions, but access requires allowlisting and Microsoft confirmation before creating test endpoints.

The change does not alter collection logic, telemetry volume, Azure Monitor Private Link Scope access modes, or Azure Monitor and Private Link data-processing charges. Its benefit is reduced private address consumption, not reduced ingestion or processing costs.

Trace GPU memory allocations back to code on AKS

Azure Kubernetes Service GPU memory profiling is in preview and measures allocations rather than GPU compute utilization. Inspektor Gadget traces CUDA allocation calls at node level without application instrumentation or sidecars, then sends results to Pyroscope for Grafana flame graphs that can be correlated with managed Prometheus metrics.

The vLLM example attributes approximately 55.1 GiB of self allocation to the function that creates key-value cache tensors. This makes the workflow useful for identifying which code holds GPU memory, rather than only observing that a device is under memory pressure.

Azure Cosmos DB walkthroughs follow an agent from prototype to production

The Code with Cosmos sessions use a travel planner to connect application development, retrieval, shared memory, and production evaluation. Together, they show how database choices affect both agent behavior and operating costs.

Start locally with AI-assisted development

Marko Hotti and Patty Chow begin with a blank folder and a product specification, then use GitHub Copilot in Visual Studio Code to build a working AI travel planner. The workflow covers data modeling, container design, queries, and local testing with the Azure Cosmos DB Linux emulator, without requiring an Azure subscription. The listed tools include Azure Cosmos DB Agent Kit and MCP Toolkit.

Combine retrieval with live operational data

The application stores operational data and vector embeddings together in Azure Cosmos DB. Full-text search handles precise queries, vector search finds semantically similar results, and hybrid search balances semantic relevance with exact-match constraints. Transactional workflows keep recommendations aligned with live inventory and booking data, connecting retrieval quality with application correctness.

Preserve context across agent handoffs

Specialized research, hotel, and restaurant agents use Azure Cosmos DB to store user preferences and shared context durably across sessions. Agents read and write that shared memory to coordinate tasks and continue work after handoffs. The demonstration references the Azure Cosmos DB Agent Memory Toolkit for implementation.

Evaluate the full production path

The production session organizes improvement around observation, evaluation, and optimization. Traces reveal end-to-end quality and performance issues, while evaluation suites replace ad-hoc manual checks with consistent assessments. Optimization spans retrieval quality, latency, token usage, memory efficiency, and Azure Cosmos DB request unit consumption, with debugging across the data, retrieval, memory, and large language model layers.

Analytics acceleration requires workload-specific checks

The analytics guidance covers both service-managed acceleration and client-side read optimization. In each case, the useful question is which workloads benefit and what compatibility or measurement constraints apply.

Fabric previews GPU-accelerated warehouse queries

GPU-accelerated query acceleration for Microsoft Fabric Data Warehouse is in public preview and requires no changes to data or SQL queries. Reported results include up to 5x faster queries at UNC Health and 3.4x faster complex workloads at single concurrency at WTW. Those are workload-specific results, not a general performance guarantee.

The same GTC Berlin announcement makes NVIDIA Cosmos Data Factory's AV Dataset Generation blueprint generally available on Azure. It supports curated, searchable, augmented, and evaluated driving datasets.

Vectored reads optimize selective Parquet access

ABFS Vectored Read combines nearby byte ranges and queues storage requests asynchronously to reduce overhead for compatible selective Parquet reads. It requires Apache Hadoop 3.3.5 or later with hadoop-azure, plus parquet-hadoop 1.12.3 or later. Spark workloads can enable it before startup with spark.hadoop.parquet.hadoop.vectored.io.enabled=true.

Azure Databricks does not currently support this feature, and Photon bypasses the ABFS driver for data access. Combining ranges can reduce requests while reading extra intervening bytes, so evaluation should compare query duration, storage transactions, bytes read, and resource utilization rather than request count alone.

Azure Databricks clarifies transaction recovery and private connectivity

Azure Databricks guidance addresses two separate production concerns: recovering from transaction conflicts and connecting serverless workloads to private resources. Both require explicit application or account configuration rather than relying on platform defaults alone.

Retry complete transactions after concurrency conflicts

Multi-table, multi-statement transactions require Unity Catalog managed Delta Lake or Apache Iceberg tables enabled for catalog commits. Non-interactive transactions use row-level conflict detection, while interactive transactions use table-level detection except for unconditional appends. When optimistic concurrency checks reject a commit, external clients must retry the entire transaction against fresh data.

The extract, transform, and load example stages, validates, and merges data atomically, using SIGNAL to roll back the block when validation fails. Transactions roll back after 48 hours, and interactive transactions also roll back after 10 minutes of inactivity. Client retry handling must therefore cover transaction boundaries, not only the statement that surfaced an error.

Private Network Gateway enters public preview

Private Network Gateway provides a managed, encrypted connection for Azure Databricks Serverless without customer-deployed gateway VMs to patch or scale. Setup requires delegating an Azure Virtual Network subnet and configuring the gateway in an account-level Network Connectivity Configuration. Workspace-level access controls, encapsulation, and network security rules enforce isolation before traffic enters the customer's subnet.

Routing can cover specified fully qualified domain names or all outbound traffic. More specific Private Link or service-endpoint paths take precedence, so teams need to account for existing connectivity when reviewing outbound routing.

Application and SQL modernization: choose the target before moving

The modernization material spans managed application hosting, Azure SQL, SQL Server on VMs, and newer database capabilities. The common decision is how much infrastructure control to retain and which migration constraints must be resolved before cutover.

Revisit managed services before renewing VM commitments

App Service removes underlying VM and operating-system maintenance, while application teams retain responsibility for data, identities, code, configuration, and network settings. Azure SQL Managed Instance and Azure SQL Database similarly provide platform-managed patching, backups, and built-in high availability. Migration guidance recommends version-controlled deployment definitions, representative testing, recovery exercises, and rehearsed cutover and rollback.

Commercial commitments need a separate review: from February 1, 2027, new eligible reservations for services covered by savings plans will no longer be exchangeable, and renewals count as new purchases. Moving a workload does not end its reservation obligation, and an Azure Reserved VM Instance does not automatically provide a platform-as-a-service benefit.

Plan Hyperscale migration and growth independently

Azure SQL Database Hyperscale separates compute, storage, and read-capacity scaling. The workload session covers serverless compute for variable demand, elastic pools for shared resources, named replicas for reads, and snapshot-based backups and continuous priming. A biomedical evidence-search demonstration combines vector and keyword search while examining growth from a pilot to a larger deployment.

The migration walkthrough uses SQL Server Management Studio assessments to identify blockers and recommend a Hyperscale target configuration. GitHub Copilot helps resolve an unused Service Broker configuration, but the BACPAC export/import approach still requires stopping writes during the export-and-import window. Source and target row counts are compared before switching the application, making validation and downtime planning explicit parts of the move.

Match Managed Instance compatibility and performance needs

The Azure SQL Managed Instance session demonstrates replication scenarios and minimal-downtime migration approaches, including discussion of existing multi-database availability groups. Performance topics include Next-gen General Purpose capabilities, configurable input/output operations per second, and flexible memory options. Update-policy choices balance earlier SQL engine features against version-aligned compatibility and portability.

Retain operating-system control with SQL Server on VMs

Azure Virtual Machines let SQL Server workloads retain control over operating-system configuration and the database environment. VM selection needs to account for memory, storage throughput, and input/output requirements, with Premium SSD v2 and Ultra Disk among the storage options discussed. Constrained vCPU configurations can optimize SQL Server licensing costs, while post-migration options include Microsoft Fabric mirroring and Microsoft Foundry.

SQL capabilities span local and cloud deployments

Anna Hoffman's Microsoft SQL overview covers database engine advances, developer tooling, and migration options across on-premises deployments, Azure, and Microsoft Fabric. Demonstrations include automatic index compaction, AI-assisted performance troubleshooting, and scalable vector search, alongside Azure Arc, Azure Local, and a Foundry-based research-agent experience.

Bob Ward's SQL Server 2025 session focuses on native vector capabilities, embedding generation, and retrieval-augmented generation patterns implemented in T-SQL. It covers local and cloud-hosted models, optimized locking, native JSON, regular expressions, change event streaming, and Data API builder. Microsoft Fabric mirroring and SQL Server on Azure Local connect those application features with deployment, security, and disaster-recovery considerations.

Use coding agents with Azure SQL development resources

The Data Exposed demonstration moves from an Azure SQL DB container to Azure SQL Database's free tier. It introduces Microsoft SQL agent skills and plugin concepts, with the Azure SQL Developer Hub and Microsoft SQL Agent Skills repository as supporting resources.

Other Azure News

Azure App Configuration's Enhanced Feature Flags public preview replaces specially formatted key-values with a dedicated resource model, independent access permissions, dedicated APIs, and validation when definitions are created or updated. Existing key-value flags remain supported, and supported provider libraries let applications retain their feature-evaluation logic.

Azure Linux 4.0 is in public preview for evaluation on Azure Virtual Machines, Virtual Machine Scale Sets, and container images, using Fedora Linux package sources while maintaining its own packages and release policies. It provides four years of full support with planned refresh windows for more impactful package changes, and carries no additional operating-system charge beyond Azure compute.

The standalone Apache Artemis JMS bridge can forward IBM MQ messages to Azure Service Bus Premium without an Artemis broker, provided its libraries use the Jakarta Messaging API namespace and Microsoft Entra ID authentication has the Azure Service Bus Data Sender role. With DUPLICATES_OK, recovery can resend messages, so stable event IDs matter for Service Bus duplicate detection and receiving applications' recognition of completed work.

An Ansys CFX study covering 168 configurations and 254 solver runs found HBv5 fastest at every tested mesh size, while HBv4 had the lowest pay-as-you-go cost per solve for the 10-million-node mesh and HBv5 for the larger tested meshes. The guidance recommends validating and pinning Message Passing Interface processes, using -part-large where meshes require 64-bit indexing, and avoiding excessive ranks on small meshes.

Network architecture guidance separates locally routable workload networks from prefixes advertised across the enterprise, using enterprise-reachable front doors to expose selected services without advertising an entire workload prefix. Enterprise-routable prefixes still need to be unique, use of 100.64.0.0/10 requires compatibility and effective-route checks, and routing must be evaluated separately from authorization.

GitHub reports that most repositories now have two of five replicas in Azure, which serves about half of Git traffic, while internal tests of per-reference write serialization showed 3-4x higher throughput. Its virtual write replica requires Azure Blob Storage durability before acknowledging writes and serves no reads, while virtual read replicas warm caches from Blob Storage without joining the write quorum.

GitHub Actions immutable subject claims include owner and repository IDs to reduce the risk of dangling Azure federated credentials granting access through reused repository names. The guidance recommends configuring those IDs in Azure before enabling immutable claims, with manual Subject Identifier edits required for enterprise claims and custom repository properties.

The remaining developer videos include a PostgreSQL teaser and a broader Azure recap. The weekly update covers networking, backup, storage, Azure Monitor, Databricks private network gateways, and AI and identity topics.

Two manufacturing clips describe agents gathering scattered telemetry, correlating signals, and ranking likely root causes with confidence levels and supporting evidence. The aim is to improve investigation and triage context, not replace the engineers who resolve incidents.

A short discussion of adaptive applications contrasts software that reasons and acts on changing context with software that only responds to user input. It points viewers to a full episode for more detail.

Also relevant this week

Dates to know

Deadlines