Scalability

Minato is designed to scale from a single game server to thousands.

Tested Limits

Metric	Tested Value	Notes
GameServers per cluster	1,000	Limited by etcd and API server
GameServers per namespace	100	Recommended for ResourceQuota
Simultaneous ActionExecutions	100	Limited by operator concurrency
Tenant namespaces	10+	With 100 GameServers each
Snapshots per server	50	Configurable via retention

Performance Characteristics

Operator

• Reconciliation rate: ~10 GameServers/second
• Memory usage: ~64MB base + ~1MB per 100 GameServers
• CPU usage: ~100m cores base + spikes during reconciliation

Control Plane

• Request latency: P50 < 50ms, P99 < 200ms
• Throughput: 1,000 requests/second
• Memory usage: ~128MB base

Agent

• gRPC connections: 1 per agent
• Memory usage: ~32MB base
• CPU usage: ~10m cores idle

Tuning Knobs

Operator Tuning

# Increase workers for faster reconciliation
args:
  - --max-concurrent-reconciles=10

# Increase cache size
resources:
  limits:
    memory: 512Mi

etcd

For large deployments (> 500 GameServers):

• Increase etcd memory: --quota-backend-bytes=8589934592
• Enable etcd defragmentation
• Use dedicated etcd nodes

API Server

• Increase max-requests-inflight
• Increase max-mutating-requests-inflight
• Enable API priority and fairness

Known Limits

1. etcd object size: 1.5MB per object
2. GameSnapshots with many entries may approach this

3. Use retention to limit entries

4. Lease-based leader election: ~15s failover

5. For faster failover, reduce lease duration

6. Webhook latency: Adds ~100ms to requests

7. For high-throughput, consider caching webhooks

8. Prometheus metrics cardinality:

9. minato_gameservers has labels: state, profile, namespace
10. Limit unique profile + namespace combinations

Scaling Checklist

• [ ] Monitor etcd memory and latency
• [ ] Monitor API server request rates
• [ ] Set ResourceQuotas per namespace
• [ ] Configure operator resource limits
• [ ] Enable horizontal pod autoscaling for control plane
• [ ] Use dedicated nodes for game servers
• [ ] Configure NetworkPolicies to limit cross-talk
• [ ] Set up log aggregation before scaling
• [ ] Test disaster recovery procedures
• [ ] Document capacity limits per cluster