Skip to main content

Region Requirements

Scope

Applies to every DynamoAI product. Cloud-agnostic; the commands that verify each requirement are in Cloud Reference.

Regions within one geography are not equivalent. The second region of a geography often carries fewer availability zones and a narrower set of GPU families than the first. Confirm each requirement below against both regions before committing to a pair. Several of them cannot be changed after provisioning.

Requirements

RequirementWhyVerify
Availability zones in both regionsThe standby provides the availability of its own region. A region without zones cannot host zone-redundant datastores or spread replicas across zonesCloud Reference
The GPU families the guardrail models require, in both regionsGuardrail serving is pinned to the GPU it was sized for. A region without that family needs a different instance type and a different card count for the same throughputCloud Reference
Quota for the standby GPU fleetCapacity a region offers is still subject to your own quota, which is set per account and per regionCloud Reference
A PostgreSQL tier that supports high availability and read replicasThe cheapest tier on most clouds supports neither, at any sizeDatastores
A dedicated MongoDB cluster tierFree and Flex clusters offer no point-in-time recovery and no multi-region optionDatastores
A global load balancer with health probingRegional withdrawal is driven by probe resultsTraffic Routing
A voting majority that survives the loss of one regionA replica set elects a primary only while a majority is reachableData Replication
Verify both regions of the pair

Availability zone support and GPU availability vary between regions of the same geography. Some regions offer no availability zones at all, and some offer a single GPU family. A pair chosen on the assumption that both regions are equivalent can pass design review and then fail to host the standby.

Decisions That Cannot Be Reversed

These are settled at provisioning time. Changing them later means rebuilding the resource.

ComponentDecisionConsequence
Node poolsAvailability zonesFixed per node pool. Changing zones requires a new pool and a workload migration
PostgreSQLCompute tierThe cheapest tier can never gain high availability or read replicas. See Cloud Reference for the tier names
PostgreSQLGeo-redundant backupCan only be enabled when the server is created
RedisHigh availabilityCan be enabled on a running instance, but never disabled. Persistence depends on it
MongoDBRegion count and voter placementTwo regions cannot survive loss of the majority region without manual reconfiguration
Object storeDistributed modeSingle-instance mode cannot be converted in place. See Object Storage

Verifying

Run the same three checks against both regions of the pair and compare the output: availability zones offered, GPU families offered, and current quota against limit. The commands differ per cloud.

A region that reports no availability zones cannot host the standby as designed.