Reliability deep dive

Live migration is not high availability.

One moves a running workload under controlled conditions. The other decides what to do after failure—when the source may be unreachable and nobody can safely ask it to stop.

Inventoryknows where a VM isMigrationmoves its running stateHAcoordinates failure recovery

Pre-copy migration

Can the network copy memory faster than the guest changes it?

Pre-copy sends memory while the VM continues running. Each round recopies pages dirtied since the previous pass. Eventually the system pauses the guest, transfers the remaining state and resumes it on the destination.

Convergence pressuredirty-page rate ÷ usable migration bandwidthIf the ratio approaches or exceeds 1, repeated copying may not converge.
  1. Round 1Copy most memoryguest keeps writing
  2. Round 2Copy dirty pagesnew pages become dirty
  3. Final pauseCPU/device state + remainderdowntime window
  4. ResumeDestination becomes activevalidate network and storage

A measurable example

Use numbers before promising “zero downtime.”

VM memoryDirty rateUsable bandwidthInterpretation
32 GiB300 MiB/s800 MiB/sLikely room to converge
32 GiB900 MiB/s800 MiB/sPre-copy may not converge

These are illustrative inputs, not a VMBook benchmark. Real measurement must include compression, protocol overhead, storage transfer, device state and the permitted downtime.

Failure timeline

HA begins where a normal migration cannot.

00:00

Host stops responding

Network loss and host failure may look identical from the controller.

00:15

Quorum decides

Surviving members determine whether they have authority to act.

00:20

Fence the source

Power or storage fencing prevents two hosts from writing the same machine state.

00:35

Recover the workload

Restart from accessible storage, then validate dependencies and service state.

Why fencing is not optional

If a disconnected host is still running while another host starts the same VM against shared writable storage, both can believe they are authoritative. That split-brain condition can corrupt data. A ping timeout is not proof that the old writer is dead.

Compatibility checklist

  • CPU baseline: destination must expose every guest-visible feature required by the running state; host-passthrough can improve performance while reducing migration compatibility.
  • Storage: shared, replicated or separately migrated with ordering and consistency understood.
  • Network: destination must preserve required VLAN, address, MTU and security policy.
  • Devices: passthrough hardware and host-specific devices can block migration.
  • Recovery: application consistency and restart behavior still need testing after the VM resumes.