Skip to content

CONCEPT Cited by 1 source

Continuous reprediction

Continuous reprediction is the online policy of re-running a predictor as new observations arrive, replacing a single commit-at-creation prediction with a stream of updated estimates. The scheduler (or other downstream consumer) consumes the latest estimate, not the initial one.

Named directly in Google Research's 2025-10-17 LAVA post: "This system uses a process we call 'continuous reprediction', which means it doesn't rely on the initial, one-time guess of a VM's lifespan made at its creation. Instead, the model constantly and automatically updates its prediction for a VM's expected remaining lifetime as the VM continues to run" (Source: sources/2025-10-17-google-solving-virtual-machine-puzzles-lava).

Why it matters

The primary motivation is recovery from early-stage misprediction. If the decision a predictor drives is expensive to reverse (VM placement being the canonical example), a single bad guess at t=0 can poison the decision until the workload actually exits. Continuous reprediction re-opens the decision whenever the estimate changes enough to justify acting.

Secondary motivations:

  • Observational evidence accumulates. At t=0, the predictor has only creation-time features (request shape, labels, user). At t=k, it has observed trajectory: actual resource usage, idle periods, lifecycle events. The later estimate is fundamentally better-informed, not just better-predicted.
  • Distribution shift within a single run. Workloads change character (burst → steady → idle → teardown). A single creation-time prediction can't reflect this; a continuously repredicted estimate can.

Mechanism

Three conceptual components:

  1. Online predictor — takes (creation-time features, observed-trajectory features so far) and emits a prediction (often a distribution — see learned lifetime distribution).
  2. Update trigger — when to re-run. Options: periodic (every N minutes), event-driven (lifecycle event, resource- usage inflection), prediction-driven (previous prediction's uncertainty hit a threshold).
  3. Consumer that acts on updated estimates — load-bearing but often under-acknowledged. If the scheduler can only use the creation-time estimate, continuous reprediction is wasted. Downstream consumers need a mechanism to react: LARS is the LAVA family's answer.

Distinct from [

deterministic simulation](<./deterministic-simulation.md>) / control loops

Control theory has online re-estimation (Kalman filters, receding-horizon MPC, online system identification); all of those are forms of continuous reprediction. The 2025-10-17 post applies the discipline to ML-based lifetime prediction in a cluster scheduler — a setting where the prediction target is discrete (does this VM still exist?), the evidence arrives in bursts, and the downstream consumer is a bin-packer, not a controller. The wiki frames it as an ML-for-systems primitive.

Contrast with single-shot ML prediction

The 2025-10-17 framing makes the contrast explicit:

Aspect Single-shot Continuous reprediction
Features used Creation-time only Creation-time + observed trajectory
Misprediction recovery Only via VM exit or migration Next prediction window
Downstream commitment Strong (one decision, long hold) Weaker (re-decidable any update)
Added infra None Prediction runner, update trigger, rescheduling hook

When it's the right shape

  • The prediction target evolves and remains measurable as the process runs.
  • The downstream decision is expensive to reverse relative to the reprediction cost.
  • A single bad prediction has asymmetric downside (cost of being wrong ≫ cost of re-running the predictor).
  • The system has a mechanism to act on updated predictions (re-scheduling, re-routing, re-allocating).

When it's the wrong shape

  • Target is fixed at t=0 and doesn't evolve with observation (e.g. one-shot classification tasks).
  • Downstream decision is cheap to re-make, so single-shot plus override-on-next-event is simpler.
  • No trajectory-observable features — the creation-time prediction is the best information available.
  • Reprediction + re-scheduling cost exceeds the misprediction cost in expectation.

Seen in

Last updated · 200 distilled / 1,178 read