Reference architecture

Reference architecture: AKS platform with GPU inference

A Kubernetes platform laid out by purpose: separate node pools for gateway, data, database and GPU inference. This mirrors how we run our own production platform.

What this architecture covers

Design decisions

  • Node pools separated by workload type, each with labels and taints
  • GPU node pool reserved for model serving
  • A replicated database on dedicated nodes with pooling
  • Registry, build runners and deployment from source control on the cluster
  • Metrics, logs and alerts for platform and workloads
What you receive
  • The architecture sized for your workloads
  • Cluster and node pool definitions as code
  • An upgrade and capacity plan

Billed at $150.00 per hour. Ask Keith for an estimate.

Diagram

The layers, top to bottom.

EntryIngressCertificatesDNSApplicationsGateway poolWorker poolAIGPU inference poolModel serversDataDatabase poolObject and file storagePlatformRegistryGitOpsMonitoringInfrastructureAKS on Azure Local
Field notes

What we check first.

Pools prevent noisy neighbors

Putting databases, gateways and GPU jobs on separate nodes keeps one from starving another.

Control plane availability

A production cluster uses a highly available control plane. We size it from the start.

Address space

Each pool, load balancer and upgrade needs addresses. The plan covers growth.

How an engagement runs

Four steps, written down.

I
Scope

A call with Keith

You describe the goal, the estate and the deadline. Keith turns it into a written scope with an hour estimate at the hourly rate.

II
Design

A written design by Gus

Gus reviews the environment and gives you a written design with each decision and its reason.

III
Build

Hands-on implementation

Gus implements in your tenant and your racks, remotely or on site, with every change logged as it is made.

IV
Hand over

Documents and a named contact

You receive the as-built record and the runbooks. Keith stays on as your account and project manager.

Questions

Asked before every engagement.

How many node pools are too many?

Each pool adds overhead. We create one where isolation or hardware differs, not for every team.

Can this start small?

Yes. The same layout works with one node per pool and grows by adding nodes.

What does the consulting cost?

Consulting is billed at $150.00 per hour. Keith sends a written hour estimate after the scoping call, and work starts when you approve it.

Rate and contacts

One rate. Two people.

Consulting is billed at $150.00 per hour. Keith runs sales, project management and your account. Gus does the architecture, the design and the implementation.

Hourly rate
$150.00

per consulting hour, for every consulting service

Sales, project management and account management

Keith

keith@apexvanguard.ai
Architecture, design and implementation

Gus

gus@gusitllc.com

Independent practice, not affiliated with Microsoft. Azure Local consulting is part of the infrastructure consulting practice of Gus IT LLC, an independent firm. It is not affiliated with, endorsed by or sponsored by Microsoft Corporation, Dell Technologies, Fortinet or any other vendor named on this site. Microsoft, Azure, Azure Local, Azure Stack HCI, Azure Arc, AKS, Azure Virtual Desktop, Hyper-V and Windows Server are trademarks of Microsoft Corporation. Dell, PowerEdge, PowerSwitch and iDRAC are trademarks of Dell Inc. Microsoft Entra, Active Directory, SQL Server, FSLogix and BitLocker are also trademarks of Microsoft Corporation. Fortinet and FortiGate are trademarks of Fortinet, Inc. VMware and vSphere are trademarks of Broadcom Inc. Kubernetes is a registered trademark of The Linux Foundation. PostgreSQL is a trademark of the PostgreSQL Community Association of Canada. Linux is the registered trademark of Linus Torvalds. All other marks belong to their respective owners and are used for description only. Product capabilities follow each vendor's documentation and change between releases; we confirm the current state during scoping.