# 专题跟踪 · 可靠性 · 2026-09-03
> 日期：2026-09-03（Asia/Shanghai）。对应卷宗 `topics/reliability.md`。SmartX 与 Arcfra 分列；RPO/RTO 保持厂商自称。

## 摘要
- VCF 9.1 仍以 vSAN stretched + 独立 Witness 为站点双活主路径；Compute-only 拉伸文档写可不配 witness。管理域须先拉伸。HA isolation address 要指到 vSAN 网。
- SmartX 榫卯 6.3（2026-03-24）：VM 级同步复制（厂商称 RPO=0）、拉伸双活门槛称降至 4 节点、SR-IOV/HCT/vGPU HA、CloudTower 管控 HA。选型文明确拉伸自动切换 vs 同步复制人工切换。
- Nutanix Metro/Sync 仍在 Ultimate/Adv Replication 墙后；NC2 AWS 上的 Metro 要用 Witness 事件驱动 NCM Playbook 改 TGW。
- HPE 9.0：拉伸与 2 节点 GFS2 共用 Distributed Worker 见证，且不能一证多用。VME 8.x HCI 布局仍是 ≥3 主机。
- NVIDIA Dynamo Snapshot（CRIU + cuda-checkpoint）与 Shadow Engine Failover 把推理恢复做成 K8s/CUDA 层，不是 hypervisor FT。vSphere 9.1 获 NVIDIA-Certified Hypervisor。ZStack ZLR 调研日未 GA。

## 原文摘录

### [Stretching vSAN Clusters in VMware Cloud Foundation](https://techdocs.broadcom.com/us/en/vmware-cis/vcf/vcf-9-0-and-later/9-1/building-your-private-cloud-infrastructure/stretching-clusters.html)
- 来源：VCF 9.1 / Last Updated July 27, 2026
> You can stretch a vSAN cluster in a workload domain across two availability zones within a region. Both availability zones must contain an equal number of hosts to ensure failover in case any of the availability zones goes down. … The default management domain vSphere cluster must be stretched before a workload domain cluster can be stretched. … vSAN Compute Only clusters support a two-node configuration (one host per availability zone). A witness host is not required for this specific topology.

### [vSAN Stretched Cluster Guide](https://www.vmware.com/docs/vsan-stretched-cluster-guide)
- 来源：VMware 文档（VCF 9.1 段）
> VMware vSAN is fully integrated and supported with vSphere HA, and should be enabled on all vSAN clusters. … Since ESXi hosts will be equally divided across both sites in a stretched cluster, it is recommended that Admission Control be set to 50% for both memory and CPU. … one must change the isolation addresses … to the vSAN network to ensure that HA is testing against the network providing data …

### [榫卯超融合 6.3 发布](https://www.smartx.com/blog/2026/03/sunmao-6-3/)
- 来源：SmartX 博客 2026-03-24
> 该版本在业内率先基于超融合架构提供同步复制容灾（RPO=0）及 RDMA 存储网络跨网卡高可用、SR-IOV 高可用等高性能业务连续性保障能力 … 虚拟机级同步复制实现 RPO=0：在已经具备拉伸集群双活能力的基础上，榫卯超融合 6.3 在国内率先提供以虚拟机为粒度的 RPO=0 原生同步复制能力。相比双活拉伸集群，同步复制无需对整个集群进行保护，可自动容忍容灾网络抖动 … 榫卯超融合 6.3 将跨站点双活集群的构建门槛由传统的 6 节点优化至 4 节点起步 … 榫卯超融合 6.3 还提供了 vGPU 高可用能力

### [同步复制与拉伸集群双活选型对比](https://www.smartx.com/blog/2026/04/sunmao-6-3-dr-q004/)
- 来源：SmartX 2026-04
> | 对比维度 | 拉伸集群双活 | 同步复制 | 保护粒度 | 整个集群 | 虚拟机级别 | RPO | RPO=0 | RPO=0 | 故障切换 | 自动切换 | 人工执行 | 网络要求 | 依赖稳定低延迟L2网络 | 支持跨L3网络，容忍一定延迟与抖动 |

### [同步复制 RPO=0 实现原理](https://www.smartx.com/blog/2026/04/sunmao-6-3-dr-q002/)
- 来源：SmartX
> 进入已同步状态后，所有写入同时作用于源端与目标端，且需在两端均完成后才返回成功，实现RPO=0。当网络抖动或短时中断时自动降级，源端业务写入不中断，网络恢复后自动同步至副本，确保最终一致性。

### [AHV Metro on NC2 on AWS](https://www.nutanix.com/blog/design-overview-ahv-metro-on-nutanix-cloud-clusters-nc2-on-aws)
- 来源：Nutanix 博客
> In NCM Intelligent Operations you need to create the playbook that will trigger Alert Policy - Automatic Failover triggered by Witness. … With communication down between the witness and AZ A and between AZ A and B the Automatic Failover triggered by Witness will be triggered executing our playbook. … If the DR site/sync destination is not reachable, automatic pause is done on source to ensure business continuity.

### [HPE VME 9 stretched / 2-node GFS2](https://community.hpe.com/t5/hpe-morpheus-vm-essentials/stretched-cluster-and-2-node-cluster-with-shared-storage-gfs2-v9/td-p/7268996)
- 来源：HPE 社区（工程师回复）
> stretched clusters and 2-node clusters do share the same basic architecture and the witness is the same Distributed Worker in both cases. … As for using the same witness for more than one cluster, this can't be done today

### [HVM clusters · VME 8.0.5](https://hpevm-docs.morpheusdata.com/en/8.0.5-vme/infrastructure/clusters/mvm.html)
- 来源：HPE VM Docs
> An HVM cluster using the hyperconverged infrastructure (HCI) Layout consists of at least three hosts. … With just one host it won’t be possible to migrate workloads between hosts or take advantage of automatic failover.

### [NVIDIA Dynamo Snapshot](https://developer.nvidia.com/blog/nvidia-dynamo-snapshot-fast-startup-for-inference-workloads-on-kubernetes/)
- 来源：NVIDIA Technical Blog
> NVIDIA Dynamo Snapshot introduces checkpoint/restore for AI inference workloads on Kubernetes … The approach combines CRIU for host state and cuda-checkpoint for GPU state, orchestrated by a privileged DaemonSet that writes artifacts to shared storage and restores them on the same or different nodes. … GMS uses the CUDA Virtual Memory Management (VMM) API to decouple large model weights from the inference worker’s process lifetime

### [Shadow Engine Failover](https://docs.nvidia.com/dynamo/dev/kubernetes-deployment/deployment-guide/shadow-engine-failover)
- 来源：NVIDIA Dynamo 文档
> This is separate from Dynamo Snapshot. … It is not a hardware fault tolerance mechanism for GPU, node, or rack loss. … It does not preserve in-flight requests, network sockets, or KV cache state.

### [VergeOS Backup & DR](https://docs.verge.io/product-guide/backup-dr/overview/)
- 来源：VergeOS Docs
> Site Syncs - Offsite Replication … Incremental sync technology minimizes bandwidth usage … Automatic failover and failback capabilities … Combine local snapshots with remote replication

## 对卷宗的影响
- 18 家对照表：FT 锁步几乎仅 VMware；ZLR 未 GA；FusionCube/FusionOne 分列且后者 RPO 自称。
- AI 失败域写入 NVIDIA 层，不把 Snapshot 写成 HCI HA。
- SmartX 6.3 是今日国产可靠性主信号；Arcfra 海外 6.3 指标不得写入中国区表。