Skip to main content

Scalable

A Scalable system grows predictably without architectural failure. Every edition of the Baseline is sized for growth before the first deployment, because re-addressing a live network means rebuilding it: the address plan reserves space for 16 regions and 8 workload stages per region, a new region is a folder-level change, a new stage is data rather than code, and the workloads scale horizontally under ceilings that bound the bill.

How each edition delivers it​

Every edition shares the shape of the address plan: the organization gets 10.0.0.0/8, each region a /12, the plat group a /13 with eight /16 stage blocks, and the core group a /14. One file holds every range, a guard script rejects a private range typed anywhere else, and the regions the code builds are the regions that have folders. A second region's block is reserved in every map.

DimensionAWS Enterprise BaselineAzure Enterprise BaselineGCP Enterprise Baseline
Address planeks_vpc_map.yaml or non_eks_vpc_map.yaml, through IPAMvnet_map.yaml or vnet_map_compact.yaml, through the ipam unitnetwork_map.yaml, through the ipam unit
Reserved regionus-east-1centralusus-east1
New stageAccount map entry, plat-account templateSubscription map entry, plat-subscription templateProject map entry, plat-project template
Node scalingKarpenter under a vCPU ceiling, Cluster AutoscalerNode Auto Provisioning under a vCPU ceilingNode auto-provisioning under CPU and memory ceilings
CI capacityOptional Spot runners that scale to zeroContainer Apps job runners that scale from zeroOptional runners on GKE Autopilot

AWS Enterprise Baseline​

Each workload stage gets a /16 with /18 private subnets sized for EKS pod density, and one /14 per region stays free for a future OU; the compact map fits the organization in a /11. Every CIDR is delivered by IPAM as typed outputs, VPC units look their allocation up by ipam_key, and the choice between the two maps is made once, before the first deployment. Regions are discovered from the region.hcl files in the tree: IPAM builds pools only for live regions and the region-restriction SCP allows every region with folders. The map reserves us-east-1 for a second region, and Transit Gateway peering between regions is built and off (enable_tgw_peering, about $36 per month per peering).

A fifth workload stage such as qa is an entry in the account map, an account.hcl, a VPC map entry, a region folder whose stack file instantiates plat-account (seven landing-zone units), a place in an isolation domain and a slug in the tag policy. Blueprints then add the stage to their common.hcl, CI matrices and GitHub Environments.

DimensionHow it grows
ComputeA baseline EKS node group per stage (dev 2 to 3 nodes, staging and prod 3 to 4) under Cluster Autoscaler, Karpenter capacity under a vCPU ceiling (32, 64, 128), and the HPA
Data processingGlue worker type and count per job (G.1X, 2 workers by default); an S3 upload invokes a Lambda that starts the Glue job, so nothing runs when nothing arrives
CI capacityOptional self-hosted runners on Spot instances, capacity-optimized with on-demand failover, that scale to zero

Karpenter chooses an instance type per pending pod and prefers Spot; Cluster Autoscaler keeps the system baseline between the node group's minimum and maximum, and the two own disjoint sets of nodes.

Azure Enterprise Baseline​

The default plan assigns region 0 (eastus2, 10.0.0.0/12) and reserves region 1 (centralus, slug cus, 10.16.0.0/12) with no folders. Four of the eight /16 stage VNets are assigned, and every stage spoke has the same eight subnets, with a /18 for AKS nodes. AKS uses Azure CNI Overlay, so pods take addresses from a per-cluster range outside the VNet and consume no VNet addresses. The compact plan, vnet_map_compact.yaml, fits the same layout in a /11; check-vnet-maps.py requires both maps to have the same regions, VNets and subnet names.

The ipam unit publishes ranges only for regions that have folders and fails the plan when folders and map disagree, and the organizations unit computes the allowed locations of the region-restriction initiative from the same folders. A new stage is an entry in plat_subscriptions, a folder whose stack file instantiates stacks/plat-subscription (eight units, from the subscription baseline to the contract vault), a block in vnet_map.yaml and a place in one isolation domain in network.hcl; the firewall policy refuses to plan a stage without a domain.

DimensionHow it grows
ComputeNode Auto Provisioning adds application nodes, spot first with on-demand fallback, under nap_cpu_limit (32, 64, 128 vCPU); the system pool scales 2 to 3 nodes in dev, 3 to 4 in staging and prod; the HPA scales pods
Data processingThe ETL trigger is a Container Apps job that scales from zero to three executions on the lake's queue; each run gets its own Databricks job cluster, at most three runs at once
CI capacityThe runners are a Container Apps job that scales from zero to five ephemeral runners on the GitHub job queue, about $0.12 per runner-hour and nothing when idle

GCP Enterprise Baseline​

network_map.yaml assigns region 0 (us-west1, 10.0.0.0/12) and reserves region 1 (us-east1, slug use1, 10.16.0.0/12), laid out the same way. Four of the eight /16 stage blocks are used, and each stage has a nodes /20 with secondary ranges pods /17 and services /20: at the Web App Blueprint's default of 64 pods per node, the pod range holds 256 nodes. Each stage also has a data /22, a Private Service Connect subnet and a Private Service Access range.

A VPC is a global resource, so a second region adds regional subnets on the same Shared VPCs and a Cloud NAT per VPC, with no new peering. The ipam unit publishes only live regions, and gcp.resourceLocations allows the home region, every region with folders and the us multi-region, so adding a region folder opens that region and nothing else does. A new stage is a plat_projects entry, a folder whose stack instantiates the plat-project template, a block in network_map.yaml, a place in one isolation domain and the stage in the units that create one resource per stage.

DimensionHow it grows
ComputeNode auto-provisioning on GKE creates nodes for pending pods under nap_cpu_limit (32, 64, 128 vCPU) and nap_memory_limit (128, 256, 512 GiB); the HPA scales pods
Data processingEach input file triggers a Dataproc Serverless batch (2 executors of 4 cores by default); the Cloud Run trigger runs at most three batches at once
CI capacityOptional Actions Runner Controller runners on GKE Autopilot (enable_github_runners, off; about $73 per month cluster fee plus pod resources while jobs run)

How you verify it​

AWS Enterprise Baseline​

  • IPAM console in the network account: the organization pool, one regional pool per live region, and the core and plat pools beneath it.
  • kubectl get nodes -L karpenter.sh/nodepool against a cluster: the baseline nodes and the Karpenter ones, labeled.

Azure Enterprise Baseline​

  • vnet_map.yaml in the Baseline repository: the assigned and reserved regions, and the free stage blocks.
  • kubectl get nodepools.karpenter.sh against a cluster: the default NodePool and its CPU limit.

GCP Enterprise Baseline​

  • network_map.yaml in the Baseline repository: the reserved us-east1 block and the four free stage blocks.
  • gcloud container clusters describe acme-usw1-prd-gke --region us-west1 --project acme-plat-prod: the auto-provisioning resource limits.

In every edition, kubectl get hpa -n <app namespace> against the prod cluster shows the 3 to 10 replica range. The landing zones page explains why stages are separate accounts, subscriptions or projects, and the Reliable page describes the same values from the failure side.