Skip to main content

Google Cloud

The reliability requirements are the same on every cloud. This page carries the Google Cloud answers.

Verified against the platform, not against a deployment

DynamoAI runs on GKE and the requirements below are the same ones the platform needs anywhere. The commands are current gcloud forms. Where a row is blank, the equivalent has not been exercised on a real deployment and is not asserted here.

Verifying A Region​

# Zones in the region
gcloud compute zones list --filter="region:( <region> )" --format="value(name)"

# GPU types offered, per zone
gcloud compute accelerator-types list --filter="zone:( <zone> )" \
  --format="table(name,zone)"

# Quota and usage for the region
gcloud compute regions describe <region> \
  --format="table(quotas.metric,quotas.limit,quotas.usage)"

Tiers That Carry High Availability​

RequirementGoogle Cloud serviceWhat satisfies it
PostgreSQL with high availabilityCloud SQL for PostgreSQLRegional availability, which maintains a standby in a second zone
Cross-region PostgreSQL standbySameA cross-region read replica, promoted manually
Redis with high availabilityMemorystore for RedisStandard tier. Basic tier has no replica and no failover
Zone-redundant block storagePersistent DiskRegional persistent disks replicate across two zones. Standard zonal disks pin a pod to one zone
Zone-redundant object storageCloud StorageMulti-region or dual-region buckets

Settings That Cannot Be Changed Later​

ComponentSettingConsequence
Node poolsZonesA zonal node pool cannot become regional. Create a regional pool and migrate
Cloud SQLZonal to regionalAvailability type can be changed on a running instance, which differs from some other clouds

Region Pairs​

Google Cloud has no designated region-pair concept. Dual-region and multi-region are storage placement options rather than a relationship between compute regions. Choose the second region on zones, accelerator availability and data residency.

Last updated on