Skip to main content

Paired Regions

Scope

Applies to every DynamoAI product, with guardrail specifics called out per page. Cloud-agnostic; see Cloud Reference for per-cloud commands.

A paired-region deployment runs the platform in two regions of one geography. One region serves production traffic. The second runs the same charts against replicated datastores and takes traffic when the first becomes unavailable.

Configure each region as described in High Availability. This section covers what the second region adds.

What Is Automatic

FailureControlAutomaticTypical recovery
Pod or containerReplica count and reschedulingYesSeconds
NodeRescheduling onto surviving nodesYesSeconds to minutes
Availability zoneTopology spread constraintsYesMinutes
Datastore node within a regionManaged service failoverYesSeconds to two minutes
Guardrail policy podRescheduling, then model loadYesSeveral minutes, up to fifteen for large models
Region becomes unhealthyLoad balancer health probeYesProbe interval multiplied by sample count
Region removed from trafficLoad balancerYesWithdrawal plus DNS propagation
Datastore promotionOperatorNoMinutes
Configuration parity checkOperatorNoMinutes
Traffic cutoverOperatorNoMinutes

Datastore promotion and the traffic cutover are operator actions. Promoting while the primary is still partially available produces writes in both regions, which then have to be reconciled by hand before the deployment can be trusted.

Section Contents

PageCovers
Region RequirementsAvailability zones, GPU families and quota, datastore tiers, load balancer
Traffic RoutingRequest path, health probe settings, regional withdrawal
Data ReplicationWhat replicates, recovery points, replica set quorum
Guardrail ServingGPU redundancy, model load time, custom policy redeployment
Monitoring And TargetsSignals to collect, recovery objectives, drills