Enterprise Metadata Management System  ·  View 16 of 22  ·  Operations

Deployment and Infrastructure

Where it runs, how it survives an availability-zone loss, and how it reaches isolated networks.

Editable source SVG draw.io All views
AWS eu-west-1 · primary · active
AWS eu-west-1 · primary · active
AZ-a
AZ-a
EKS node group
services · workers
EKS node group...
Aurora writer
PostgreSQL 16
Aurora writer...
OpenSearch data
OpenSearch data
AZ-b
AZ-b
EKS node group
EKS node group
Aurora reader
failover target
Aurora reader...
OpenSearch data
OpenSearch data
AZ-c
AZ-c
EKS node group
EKS node group
Neo4j cluster
3 core members
Neo4j cluster...
MSK brokers
Kafka
MSK brokers...
AWS eu-central-1 · warm standby
AWS eu-central-1 · warm standby
Standby estate
Standby estate
EKS scaled to zero
IaC identical
EKS scaled to zero...
Aurora global replica
RPO under 1 min
Aurora global replica...
S3 cross-region copy
S3 cross-region copy
Network-isolated estates
Network-isolated estates
Restricted VPC or data centre
Restricted VPC or data centre
Ingestion agent
outbound 443 only
Ingestion agent...
Private sources
no inbound route
Private sources...
Route 53 + WAF
health-checked
Route 53 + WAF...
Entra ID
Entra ID
Backup vault
PITR 35 days
Backup vault...
HTTPS 443
HTTPS 443
global replication
global replication
mTLS outbound
mTLS outbound
read-only
read-only
continuous backup
continuous backup
Deployment and Infrastructure
Deployment and Infrastructure
Application we own
Application we own
Data store
Data store
Queue / topic
Queue / topic
Interface / broker
Interface / broker
External / third party
External / third party
Security / platform
Security / platform
synchronous
synchronous
event / async
event / async
batch
batch
Warm standby, not active-active: metadata reads tolerate a 30-minute RTO, and dual-region write consistency for a versioned store is not worth its cost here.
Warm standby, not active-active: metadata reads tolerate a 30-minute RTO, and dual-region write consistency for a versioned store is not worth its cost here.
v 1.0 · owner Platform Engineering · date 2026-08
v 1.0 · owner Platform Engineering · date 2026-08
Text is not SVG - cannot display

Decisions

  • Warm standby rather than active-active. Metadata reads tolerate a 30-minute RTO, and dual-region write consistency for a versioned store is not worth its cost or its failure modes.
  • The ingestion agent runs inside restricted networks and connects outbound only. No inbound route into a source estate is ever required.
  • Three availability zones in the primary region; every stateful component has a quorum across them.

Numbers

  • 99.9% read-path availability, RTO 30 min, RPO 5 min.
  • Aurora global replication lag under 1 minute to the standby region.
  • Steady state roughly 40 vCPU and 160 GB across the service tier; the graph and index dominate cost, not compute.

Data residency

  • Both regions are in the EU. A second residency zone would be a separate deployment with its own stores, not a shared multi-region cluster — metadata carries column names and classifications and is treated as regulated content (view 20).