Private · High Availability

Eliminate single points of failure. Consider it across the application layer, the data layer, and the model layer simultaneously.

Three-Layer HA Checklist

Application layer
☑ Multiple Evose Web/API replicas (≥ 3)
☑ Multiple Worker replicas (≥ 2)
☑ Active-active Ingress / load balancer

Data layer
☑ MySQL primary-replica (synchronous replication) + auto failover
☑ PostgreSQL primary-replica + pgvector replication
☑ Redis primary-backup / sentinel / cluster
☑ Object storage cross-region replication

Model layer
☑ Multi-LLM-vendor integration (Failover)
☑ Multi-instance self-hosted LLM (Round Robin)
☑ Multi-instance Embedding / Reranking

Application Layer

Recommended K8s configuration:

Component	Replicas	PDB (PodDisruptionBudget)
evose-api	≥ 3	minAvailable: 2
evose-worker	≥ 2	minAvailable: 1
evose-web	≥ 2	minAvailable: 1

Data Layer

MySQL

Option	Best for
Primary-replica + manual failover	Small-mid scale
MGR (Group Replication)	Recommended, production
MySQL InnoDB Cluster	Large scale, mature ops needed
Aliyun RDS / cloud DB	Managed — least overhead

PostgreSQL + pgvector

Option	Best for
Streaming replication + Patroni	Recommended
Citus sharding	Large-scale knowledge bases
Aliyun PolarDB / cloud DB	Managed

pgvector primary-replica

Vector data does sync via streaming replication, but the vector index (HNSW) needs to be rebuilt on the replica. Use replicas for read; for large-scale, consider Citus.

Redis

Option	Best for
Primary-backup + sentinel	Default
Redis Cluster	Large scale / cross-DC

Model-Layer HA

Multi-Vendor Failover

Interface platform registers the same model with multiple vendors:

gpt-4-turbo
├─ OpenAI US-West       → priority 1
├─ Azure China          → priority 2 (used when primary is down)
└─ Self-hosted OpenAI-compatible → priority 3 (used when all SaaS down)

Self-Hosted Multi-Instance

Model deployment:

qwen-max:
  ├─ Instance 1: 10.0.0.5:8000  (replicas 2)
  ├─ Instance 2: 10.0.0.6:8000  (replicas 2)
  └─ Instance 3: 10.0.0.7:8000  (replicas 2)
Strategy: Round Robin + health checks

Cross-DC / Multi-Region

Advanced. Recommended:

Primary DC                       DR DC
─────                            ──────
Full Evose stack + DB primary    Full Evose stack + DB replica
        │                              ↑
        └─── async DB cross-region replication ───┘

DNS switch / GLB swaps automatically within RPO/RTO

Cross-region DR RPO

Async replication has data-loss risk. Strict-compliance scenarios (finance / healthcare) need to evaluate sync replication and its latency cost.

Failure Drills

Drill	Frequency	Target RTO
Single API Pod failure	Monthly	< 30s
Primary MySQL failure	Quarterly	< 5min
Single LLM vendor failure	Monthly	< 10s
Whole DC failure	Semi-annually	< 30min

Next Steps

Upgrade procedure → Upgrade
6-layer defense → Security

Private · High Availability

On this page