load-aware-hedging · research · also LAEDGE, LÆDGE

Load-Aware Hedge (LAEDGE)

LAEDGE is a per-shard scheduling policy that sends redundant queries to idle replicas only when no unserved query can use them.

At dispatch, queued first executions take priority over copies; an idle replica otherwise takes the oldest unfinished, unreplicated query. Plain LAEDGE does not cancel running copies. Cleanup cancellation stops a losing copy after completion; preemptive cancellation frees a replica for a new query. [1]

The dispatch-time rule does not guarantee that a running copy never delays later arrivals. Preemptively cancelling the newest copy can discard the one that would finish first. [1]

Relations

related
Hedged request, Tied request, Adaptive replica selection, Tail latency

Limitations

  • The evaluation concerns in-memory, CPU-bound search; it does not establish a storage-device speedup. A longer scheduler-to-replica round trip can reduce throughput. [1]

Citations

  1. Primorac, Argyraki and Bugnion. When to Hedge in Interactive Services. NSDI 2021, pp. 373-387; sections 4-6, Algorithm 1. (accessed 2026-11-13)

Further reading

  1. Hesela. Load-aware hedging and the dispatch-time boundary. (accessed 2026-11-13)
Sources & caveats
  • Full proceedings paper checked, especially Algorithm 1 and sections 4, 5 and 6; simulation and cloud implementation are separate evidence.
  • No reproduction or hardware experiment; publication year is not a deployment claim.

JSON: /data/terms/load-aware-hedging.json

Published by Hesela team · Updated

Report a correction · Editorial standards · Reuse guidance

Cite this page
Hesela team. Load-Aware Hedge (LAEDGE). Hesela. 2026-11-13. https://hesela.dev/terms/load-aware-hedging/