A virtual machine can survive a failed ESXi host yet still be unavailable after a storage outage, an application crash or a regional disaster. Those events require different recovery mechanisms. VMware vSphere High Availability (HA), vSphere Fault Tolerance (FT) and disaster recovery (DR) are often discussed together, but treating them as interchangeable—or assuming every business needs all three—leads to expensive designs with uncovered failure modes. The useful starting point is a clear description of what might fail, how long the service may be unavailable, and how much data could be lost.
Start with the failure, not the availability feature
Consider three incidents affecting the same ordering application. First, one ESXi host stops running while the application virtual machine’s disks remain accessible to the cluster. Second, the host fails during a financial transaction that cannot tolerate even a guest restart. Third, a storage system or entire site becomes unavailable. Restarting the VM on another host may address the first event. A suitably supported and correctly configured FT-protected VM may reduce the interruption from the second event when the underlying failure is the host. Neither mechanism alone restores a usable database from a lost site.
Document the business targets before selecting a VMware product. RTO and RPO answer different questions: the maximum acceptable time to restore a service, and the maximum acceptable age of recovered data. They are targets, not automatic promises of any particular deployment mode. Availability requirements should also state the tolerated failure scope, transaction-consistency requirements, application dependencies and recovery ownership. A two-minute objective means little if the identity provider, DNS service or database on which the application depends is not part of the recovery design.
Separate planned maintenance from unplanned failure. Live migration and coordinated application maintenance may avoid a restart during planned host work, but they do not prove that the environment can recover from an unexpected power loss. Likewise, duplicating hosts at one location does not demonstrate protection from a building outage. A design is credible only when the tested failure matches the one the organization is promising to withstand.
vSphere HA restarts workloads after eligible failures
vSphere HA is a cluster-level recovery capability. When a protected ESXi host fails, HA can restart affected virtual machines on another compatible host with sufficient resources and access to the necessary virtual machine files. This is restart-based recovery: guest operating systems and applications need time to boot and resume service. Broadcom explicitly distinguishes HA’s restart behavior from the continuity model of Fault Tolerance. An HA dashboard showing that protection is enabled is therefore not evidence of zero application downtime.
Host monitoring depends on network communication and datastore heartbeats. Network isolation response settings influence whether virtual machines are shut down or powered off before a restart elsewhere; storage-path problems may require specific VM Component Protection behavior. VM Monitoring and Application Monitoring are separate options with their own heartbeat and restart conditions. An operator should not infer that every guest or application hang causes automatic recovery unless the relevant monitoring is installed, enabled and tested.
Admission control and capacity are operational requirements. If every healthy host is already near saturation, restarting a failed host’s workload may fail or produce severe performance degradation. Check available memory, CPU reservations, storage visibility, compatible host configuration, network access and dependency startup order under the loss of the largest anticipated host. HA does not automatically return a restarted VM to its original host after repair; Distributed Resource Scheduler or a planned administrative action may subsequently rebalance placement. A restart can also occur without a clean application shutdown, so databases and transactional services still need crash-recovery validation.
Fault Tolerance targets continuous VM execution during host failure
vSphere FT maintains a running secondary VM on another host and synchronizes execution state with the primary. For a supported VM experiencing a qualifying primary-host failure, the secondary can take over with little or no noticeable guest interruption. This is why FT differs fundamentally from HA, which boots a replacement copy after failure. But the boundary matters: Broadcom documents host-failure continuity, not universal immunity to defective application code, data corruption, lost external services or an unavailable recovery site.
If an application writes the wrong value, the synchronized secondary is not an independent clean copy of the database. If a dependency becomes unreachable from both hosts, a surviving VM may continue executing while its users still receive errors. FT also introduces operational cost: a second running VM consumes resources, state synchronization depends on reliable networks, and write-intensive or memory-churn-heavy workloads can experience performance constraints. Supported CPU, virtual hardware, VM configuration, licensing and vSphere version must be checked against current Broadcom documentation rather than copied from an old exam summary.
FT is configured per VM, not enabled indiscriminately for an entire business. Compatibility restrictions can affect snapshot usage and other VM operations, while FT failover itself should be tested in a controlled nonproduction environment. A realistic validation records both the application-side transaction outcome and the platform-side event. If a business needs continuous availability from application bugs as well as physical-host loss, it also needs application-level redundancy, resilient data handling and independent monitoring. FT is one component of that design, not a replacement for it.
Disaster recovery covers loss that a local cluster cannot contain
DR plans for events that remove or compromise a larger failure domain: a whole site, shared storage platform, major network dependency or critical data set. Recovery usually involves a separate protected copy of data, enough compute and networking to run the workload elsewhere, a coordinated sequence for dependent applications, and tested procedures for redirecting users. Replication can shorten the age of recovered data, but it can also replicate unwanted changes or corruption; independent, recoverable backups are therefore important even in replicated environments.
VMware Live Site Recovery, previously called Site Recovery Manager, has provided recovery-plan orchestration and testing for supported configurations. Exact product packaging and naming have changed over time, so a deployment plan should verify the current release, supported replication integration and entitlement with Broadcom. A recovery plan is only as complete as its dependencies: credentials, certificate services, DNS, IP changes, database consistency, security rules and external integrations may all require ordered restoration. A failover button does not make an untested application architecture site-resilient.
RPO depends on replication mode, lag and consistent recovery points; RTO includes detection, declaration, technical recovery, validation and user cutover. A plan that achieves a fast VM power-on but takes hours to reconnect identity or payment systems has missed the real service objective. Test the recovery path with representative data, confirm that staff can execute it without the original site’s management plane, and record when writes may resume safely. In regulated or high-integrity systems, the organization must also know which recovered transactions require reconciliation.
Make the choice from a concrete workload, not a generic ranking
A noncritical internal reporting service might accept a VM restart after a host failure and a slower recovery from backups after a site disaster. vSphere HA with adequate spare capacity and a tested backup or DR procedure may be a proportionate design. An order-processing component with a very low interruption budget could justify a compatible FT configuration for host failure, but it may still require a clustered database or replicated application tier to protect against software and data failures. Neither scenario establishes that every VM should use FT.
For an organization with two facilities, local HA can reduce individual host outage duration while DR addresses loss of the primary facility. These layers answer distinct failure questions and may coexist, but adding FT to every VM would introduce cost and complexity without necessarily improving the most important recovery objective. Decide based on measured business consequence, recovery time, tolerated data loss, service dependencies and the cost of maintaining reliable tests. A smaller, fully tested design is more defensible than an impressive diagram whose failover path nobody has executed.
When discussing options with nontechnical stakeholders, use plain outcomes instead of feature labels. Say, for example: “If one host fails, the VM will be restarted on a surviving host and the application may be unavailable during restart.” Or: “For this supported FT VM, a qualifying host failure should cause little service interruption, but application failures and site-wide failures require other controls.” Or: “After complete site loss, our tested recovery plan targets a specified RTO and RPO.” These statements can be verified and avoid the misleading promise of universal zero downtime.
Test availability at the application boundary
Create a failure-test matrix listing the event, protection mechanism, expected outcome, owner and evidence. Include host power loss, management-network isolation, storage-path interruption, guest operating-system hang, application process crash, database corruption and site unavailability. Not all scenarios should be simulated in production; use supported test controls and a representative staging environment. For each test, monitor completed user transactions as well as hypervisor events. A VM may show as running while its application remains unusable.
For HA, record detection and restart intervals, guest boot time, dependent-service readiness and capacity on surviving hosts. For FT, record supported configuration, secondary state, application behavior and any resource or network overhead. For DR, measure the age of recovered data, recovery-order dependencies, cutover time and successful user verification. Keep the tests repeatable, and distinguish observed results from untested expectations. If the only evidence is “the cluster is green,” the organization does not yet know its real recovery capability.
Finally, do not treat hypervisor snapshots as independent backup copies. Broadcom warns that snapshots are not a backup strategy for a VM’s virtual disks, and a storage failure can affect both a VM and its local snapshots. Retained offline backups, application-consistent recovery points and verified restoration procedures serve a different purpose. The operational discipline is straightforward: name the failure, select the minimum controls that genuinely address it, validate the business outcome and revise the plan whenever infrastructure or applications change.