← all builds

From-Scratch Build · 05 · Data & Autoscaling

Digital Twin Scalability Dataset

What happens to an edge service when you keep adding robots? This is the dataset that answers it — and the machine-learning angle on top: teach a model to predict when a digital-twin service is about to break its SLA, so it can scale before it does.

DatasetEdgeSLA management Random ForestAutoscaling

What it is

A stress test, recorded

This is the sibling project to the Robot-Arm Digital Twin. Where that one was the system, this is the measurement: a dataset captured while progressively spinning up more robot-manipulator instances against the same edge-based digital-twin service, watching how resource use and responsiveness degrade as load climbs.

It comes in three flavours that differ only in how aggressively new robots arrive — and a labelled version exists to train a classifier, which is where the project stops being "a CSV" and becomes a question about intelligent autoscaling.

60s
Micro dataset
+1 robot / minute
300s
Small dataset
+1 robot / 5 min
3600s
Big dataset
+1 robot / hour

The stack

From telemetry to decision

subject

Edge Digital Twin

The robot-arm twin service under test — the thing being scaled and measured.

data

Time-series telemetry

Resource and performance metrics logged as robot instances are added at fixed intervals.

ML

Random Forest

A classifier trained on the labelled big dataset — an ensemble of decision trees that votes on the system's state.

objective

SLA management

Service-Level Agreements set the threshold; the model predicts breaches so scaling can act in time.

action

Autoscaling

The payoff: scale the service up/down based on a prediction, not a lagging alarm.

context

AIML-as-a-Service

The same prediction can be offered as a network service — autoscaling intelligence delivered on demand rather than hand-tuned.

How it works

The experiment loop

  1. Define the arrival rate

    Choose micro / small / big — i.e. how fast new robot instances join the load.

  2. Scale & record

    Spin up instances on schedule, logging resource and timing metrics throughout the run.

  3. Label the states

    Tag samples (e.g. SLA-met vs. at-risk) to create the supervised target for the big dataset.

  4. Train the classifier

    Fit a Random Forest to map live telemetry to a predicted SLA state.

  5. Predict & scale

    In deployment, the model flags imminent breaches so the orchestrator scales ahead of failure.

Reflection

What rebuilding it taught me