The scheduler

Scale that
respects back-pressure.

The Omega scheduler sees two axes — the substrate (Gaia / Biome / Habitat) and the capability (Standard / Frontier) — and a queue depth per edge. It scales when the queue grows, places replicas to maximize spread, and stops when back-pressure clears. No CPU-percent guessing; no surprise bills.

[axis 01]
PlacementSubstrate
[axis 02]
CapabilityProfile
[signal]
edge queue depth · P95 latency
[strategy]
queue-aware · predictive
At a glance

The properties
that matter.

< 200 ms
scale-up reaction

From queue-fill signal to a new replica answering. Standard tier on Gaia.

0
CPU-percent triggers

We don't autoscale on CPU%. Queue depth and tail latency are the signals that match the work.

replica ceiling

Limited only by your tenant quota and the substrate capacity you declare.

AZ-aware
replica spread

Replicas spread across hosts and zones automatically. Weight by latency, by host, by region.

In the manifest

How you
describe it.

scaling.zgraph.toml
[brane.score]
substrate   = "gaia"
capability  = "standard"
image       = "omega/score:1.4"
// queue-aware autoscale, predictive warmup

// queue-aware autoscale, predictive warmup

In operation

What it
looks like running.

$ω scaling status --brane score
brane.score current 128 / 256 (target: 138, P95-aware) queue 21 / 32 target (8% headroom) P95 118 ms predictive +14 in 4m (warmup) …
$ω scaling history score --since 1h
13:00 64 13:15 98 ↑ queue=42 → spawn 24 13:33 118 ↑ queue=51 → spawn 20 (P95=212ms) 13:42 128 → steady 14:00 108 ↓ queue=8 → drain 20
signals
queue depth · P95 latency · cost
placement
two-axis · substrate-aware spread
back-pressure
first-class · upstream sees not-ready
predictive
rolling window · diurnal-aware
cost ceiling
hard wall · scaler refuses to exceed
scale-down
drain queues · graceful · zero loss
By design

What this
surface does.

Scale on the work, not the metric

Edge queue depth is the truth. CPU% lies under burst. P95 lies during cold start. Queue depth is the load.
QUEUE-AWARE

Learns the curve, warms ahead

Most workloads have shape — diurnal, weekly, around marketing pushes. The scheduler learns the shape and pre-warms replicas before the wave arrives.
PREDICTIVE

Substrate-aware placement

Replicas spread across hosts, AZs, and substrates. Lose a Gaia AZ? Branes migrate to other AZs (or to Biome) according to manifest preferences.
SPREAD

A hard wall on hourly spend

cost_ceiling is checked at every scale decision. The scheduler refuses to spawn replicas that would push you past the wall — even if the queue keeps growing.
COST