Multimodal RAG Platform · View 11 of 16
Deployment & Infrastructure
What runs where, across two availability zones and a warm standby region.
Copy
PNG
PDF
⋯
Editable source
SVG
draw.io
All views
Primary region — active
Primary region — active
Availability Zone A
Availability Zone A
Ingress Controller
Ingress Controller
General Node Pool
8 vCPU x 3
General Node Pool...
GPU Node Pool
1 x L4, rerank + VLM
GPU Node Pool...
Egress Proxy
allow-listed
Egress Proxy...
Availability Zone B
Availability Zone B
Ingress Controller
Ingress Controller
General Node Pool
8 vCPU x 3
General Node Pool...
GPU Node Pool
1 x L4, spot burst
GPU Node Pool...
Egress Proxy
standby
Egress Proxy...
Managed services — multi-AZ
Managed services — multi-AZ
Metadata DB
HA pair, PITR
Metadata DB...
Vector Service
3 nodes, RF 2
Vector Service...
Search Cluster
3 data nodes
Search Cluster...
Object Store
versioned, CMK
Object Store...
Cache
primary + replica
Cache...
Work Queue
DLQ per stage
Work Queue...
DR region — warm standby, RPO 15 min / RTO 4 h
DR region — warm standby, RPO 15 min / RTO 4 h
Object Replica
cross-region copy
Object Replica...
DB Read Replica
streaming
DB Read Replica...
Pilot Cluster
0 replicas until failover
Pilot Cluster...
Index Rebuild Job
from derived zone
Index Rebuild Job...
CDN / Edge
static assets
CDN / Edge...
Claude API
regional endpoint
Claude API...
replication, 15 min
replication, 15 min
streaming replica
streaming replica
rebuild on failover
rebuild on failover
Deployment & Infrastructure
Deployment & Infrastructure
Interface / broker
Interface / broker
Application we own
Application we own
Security / platform
Security / platform
Data store
Data store
Queue / topic
Queue / topic
External / third party
External / third party
event / async
event / async
batch
batch
Vector and lexical indexes are not replicated. They are rebuilt from the derived zone, which is cheaper than paying to mirror them.
Vector and lexical indexes are not replicated. They are rebuilt from the derived zone, which is cheaper than paying to mirror them.
v 1.0 · owner Platform Engineering
v 1.0 · owner Platform Engineering
Text is not SVG - cannot display
Failure domains
Two availability zones, each with ingress, general and GPU pools
Managed data services span both zones
Zone loss costs capacity, not availability
Disaster recovery
RPO 15 minutes, RTO 4 hours
Object store replicated cross-region; database streams a read replica
Pilot cluster runs at zero replicas until failover
A deliberate omission
Vector and lexical indexes are not replicated
They are rebuilt from the derived zone on failover
Cheaper than mirroring, and within the 4-hour RTO
