1. Don’t panic, but know where the spare capacity is

Douglas Adams’ The Hitchhiker’s Guide to the Galaxy returns as the recalled work because its advice applies to schedulers too: know what you have, keep a backup plan, and do not rely on a single point of failure. A self-migrating compute system needs a few simple, reliable primitives that the cells can use to advertise health, discover capacity, and move workloads.

This entry lists what the desktop would have to provide for autonomous compute migration to work.

2. Health monitoring

The scheduler can only react to problems it can see:

  • Sensor telemetry: temperature, voltage, current, error rates, memory correctable errors, and radiation events.
  • Application metrics: latency, throughput, and checkpoint success.
  • Trend detection: moving averages and thresholds that distinguish degradation from noise.
  • Predictive hints: models that estimate time-to-failure from historical patterns.

3. Workload portability

A workload can only migrate if it is not tied to one piece of hardware:

  • Containerized or virtualized payloads with defined resource requirements.
  • Stateless compute where persistent data lives in the federated storage layer.
  • Checkpoint and resume for long-running tasks that cannot restart from zero.
  • Hardware abstraction so a payload does not depend on a specific GPU, FPGA, or sensor interface.

4. Target selection

The scheduler needs a ranked list of eligible cells:

  • Resource availability: CPU, memory, power, and thermal budget at the target.
  • Affinity and anti-affinity rules: some workloads must stay near their data or away from each other.
  • Network topology: migration is cheaper if the target is on the same switch or rack segment.
  • Security and policy constraints: not every payload can run on every cell.

5. Safe execution

Migration must not make things worse:

  • Pre-flight checks: verify target health and resources before starting the move.
  • Atomic handoff: ensure the workload is never running in two places at once unless explicitly designed for it.
  • Rollback: if the new cell fails, return the workload to the original or another known-good cell.
  • Audit log: every migration is logged with cause, source, target, and outcome.

6. Policy and limits

A scheduler that moves workloads must still obey rules:

  • Migration windows: some workloads move only during low-activity periods.
  • Rate limits: prevent a cascade of migrations from overwhelming the network.
  • Human approval: critical or security-sensitive moves may require ground confirmation.
  • Quarantine: a misbehaving cell can be excluded from receiving new workloads.

7. What this changes

  • Autonomous compute migration requires investment in health monitoring, workload abstraction, target selection, and safe execution.
  • Many of these provisions are worth making even for ground-commanded workload assignment.
  • The next entry will decide whether the desktop should design for this capability.