JournalDAY 27 / X

FIELD NOTE / X

Backfill protects an estimate, not your contract.

The complete written thought and the evidence behind it. The video edition will follow its public release.

Journal September 25, 2026 · X target October 24, 2026

Backfill protects an estimate, not your contract.

Video caption

Backfill protects an estimate, not your contract. Idle space fills while the reserved customer waits. Keep the efficiency gain only when the SLA holds. #EricFieldNotes

Full written post / accessibility read

Slurm backfill starts lower-priority work only when it should not delay the expected start of higher-priority work. That expected time depends on time limits and current state. A customer promise is a separate clock. If the two disagree, a busy fleet can still deliver a late job.

The provider can show higher utilization from short jobs. The reserved workload may still start after its contract time because the initial estimate, topology, failure or reservation policy was wrong. The small job is visible in the timeline but not automatically causal.

For each reservation, store the external deadline, Slurm's start estimate and changes to it, time limits, topology fit and actual allocation. In a test queue, vary job duration estimates and failures, then inspect which condition produced each miss.

Use backfill to improve utilization, but audit external start commitments separately and revise estimates or reserve policy when they fail. Do this because an algorithm can honor its expected-start rule while the product still fails its buyer.

#EricFieldNotes

Four-beat scene transcript

1. Backfill protects an estimate, not your contract.

Slurm backfill starts lower-priority work only when it should not delay the expected start of higher-priority work. That expected time depends on time limits and current state. A customer promise is a separate clock. If the two disagree, a busy fleet can still deliver a late job.

Visual: Correct scheduler behavior can still miss an external start promise.

2. The dashboard rewards the wrong success.

The provider can show higher utilization from short jobs. The reserved workload may still start after its contract time because the initial estimate, topology, failure or reservation policy was wrong. The small job is visible in the timeline but not automatically causal.

Visual: Idle space fills while the reserved customer waits.

3. Compare three clocks.

For each reservation, store the external deadline, Slurm's start estimate and changes to it, time limits, topology fit and actual allocation. In a test queue, vary job duration estimates and failures, then inspect which condition produced each miss.

Visual: Promised, expected at admission, and actual start.

4. Backfill with a promise ledger.

Use backfill to improve utilization, but audit external start commitments separately and revise estimates or reserve policy when they fail. Do this because an algorithm can honor its expected-start rule while the product still fails its buyer.

Visual: Keep the efficiency gain only when the SLA holds.

Research and claim limits

The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.

More notes from the work ↗