Working AtlasControl loops & scaling

Autoscaler Arena

Hand-draw a load spike, then try to keep the cluster from oscillating.

TryRun Sawtooth with stabilization on. Turn it off and select Sawtooth again.

Watch forStabilization delays scale-down; compare replicas under the same changing load.

Jump to the model

Separate teaching model

The linked story describes related work; the scenarios and results here are synthetic.

Interactive model

1 · Choose a load profile

Choose a starting pattern. Next, shape the curve and compare how the scaling rule responds.

2 · Compare scale-down behavior

3 · Shape the workload

Drag on the curve or use the precise controls below it. The amber marker is the point you are editing; the coral dot is live playback.

Edit a load point

Selected time point02:03

Requests per second at selected point100

Editing a point changes the workload without moving playback. Replay run starts the same experiment from the beginning.

Warming up…

Wait for the first HPA sync tick before interpreting the response. When reduced motion is enabled, use Step model to advance the run.

4 · Watch the response

Compare the applied replicas, derived latency, and warming pods as the selected load runs.

Applied replicas
Desired (ideal) replicas
Latency
Latency threshold
Warmed-up pod
Pod warming up

Thrash count

—

Replica-count direction reversals

Time within threshold

—

Share of simulated time below the configured latency threshold

Pod-minutes

—

Replica x minutes (cost proxy)

How the model decides

This simplified HPA-inspired teaching model uses a 15-second tick, a ±10% no-action band, and a scale-down lookback that favors the highest recent recommendation. Its new-pod warm-up and latency curve are synthetic; Kubernetes HPA has additional metric and scale-rate rules that this model does not reproduce.

Preparing the interactive experiment…

Skill practicedTune feedback loops and scale stabilization

Connection to my workNeuralEye’s distributed inference makes capacity planning relevant; this synthetic model explores delayed scaling feedback.

Related project:NeuralEye project story →

Understand the result

A simplified HPA-inspired autoscaler watches a hand-drawn or preset synthetic load curve. Disable scale-down stabilization and oscillating load can make replicas flap; restore the window to compare the same curve under a smoothed decision rule. This is a teaching model, not a latency or cost forecast.

Controls

Load curve editor: Drag on the chart to draw requests per second over the simulated four-hour window, or select and edit individual points with the controls below it.
Selected time point: Chooses one of 48 samples across the four-hour load curve. The amber marker identifies the point being edited; the separate blue playhead shows the live run position and continues moving when you edit a point.
Requests per second at selected point: Sets the load at the selected time point and switches the curve to Custom.
Diurnal: Starts a smooth day/night cycle. Begin here, then edit a sample on the chart if you want a custom workload.
Flash crowd: Starts a sudden traffic spike. Compare its scale-up response with the other profiles.
Sawtooth: Starts a repeating ramp that makes scale-down oscillations easier to compare.
Stabilization window enabled: Run the same profile with this on and off. On: hold the highest recent recommendation for five minutes before scaling down. Off: apply scale-down immediately.
Replay run: Restarts the same deterministic scaling run from the beginning with the current load curve and stabilization setting.
Step model: When reduced motion is enabled, advances the simulation in four paced frames while automatic animation stays paused.

What to watch for

With stabilization ON, compare scale-down behavior as the preset load oscillates; recent higher recommendations can keep the applied replica count from dropping immediately.
Turn it OFF and select the same preset again — immediate scale-downs can increase the model's Thrash count.
After a sharp load spike, amber replicas are still warming up in the model; derived latency can remain above its configured threshold after scale-up has been recommended.

Why it happens

The model ignores metric changes inside a tolerance band. For scale-down, it selects the highest recommendation in its recent five-minute history instead of applying the newest lower value immediately. When an older high recommendation ages out, the fleet can scale down; load does not have to remain continuously low for the whole window. Removing the lookback makes reductions faster, which can create more reversals when load rises again.

How this idea relates to the linked work

NeuralEye processes video through model and verification stages, with compute efficiency as a real system concern. This synthetic control loop helps me reason about capacity signals, scaling delay, and the trade-off between spare workers and derived latency. The workload and scaling policy are illustrative, not NeuralEye telemetry or autoscaler settings.

See the broader engineering brief