Read-only review demo · Published historical evidence

SpotAtlas
Multi-cloud research broker

Explore the historical AWS/Azure evaluation and suggest improvements. This static page works without Python, a backend, cloud credentials or analytics.

Not a live cloud product. These are March/April 2026 archived observations, not current offers, availability guarantees or survival estimates. No VM can be launched here. Public production readiness is not established.

AWS

19,152

published observations · 84 pools · 45 types

2026-03-25T10:00:07.999467373+00:00
to 2026-04-01T12:59:29.946501017+00:00

Azure

81,288

published observations · 79 pools · 42 types

2026-03-25T10:00:04.075802326+00:00
to 2026-04-01T12:59:44.694130898+00:00

Explore forecast horizon

15-minute historical evaluation

Model selected on a separate chronological selection period: Random forest.

Lower Brier score is better. Final-test metrics are not used to select the model.

Historical capacity-acceptance prediction, not VM survival
Method	Selection Brier	Final-test Brier
Historical pool mean	0.034507	0.034552
Persistence	0.004809	0.005954
Rolling 1h	0.003703	0.004374
Logistic regression	0.003855	0.004277
Random forest · selected	0.003631	0.004057
Gradient boosting	0.004466	0.004487

60-minute historical evaluation

Model selected on a separate chronological selection period: Random forest.

Lower Brier score is better. Final-test metrics are not used to select the model.

Historical capacity-acceptance prediction, not VM survival
Method	Selection Brier	Final-test Brier
Historical pool mean	0.034864	0.035146
Persistence	0.007591	0.007325
Rolling 1h	0.005530	0.004737
Logistic regression	0.005904	0.004980
Random forest · selected	0.005199	0.004758
Gradient boosting	0.005555	0.005426

Browser-only illustrative walkthrough · No backend execution

Submit once. Checkpoint. Recover.

The private local job engine now supports durable SQLite job records, worker leases, verified checkpoints, bounded retries and cancellation. A local fixed hash workload was tested: interrupted at 200 of 1,000 steps, then resumed by a second worker to completion.

The controls below replay a scripted example only. They do not submit a real job, run the Python worker, access the private database or launch a cloud VM.

Walkthrough ready · no job submitted

Local engine evidence is labelled local_checkpoint_demo, not provider measurements. Container checkpointing, cloud execution and cross-cloud migration are not implemented by this walkthrough.

What is implemented?

Historical processing/evaluation, evidence-gated local ranking, optional read-only provider metadata adapters, reviewed AWS experiment tooling, scoped cleanup and billing tools.

Also implemented locally: persistent job submission, idempotency, worker fencing and fixed-workload checkpoint/resume. These private tools are not exposed through this public static page.

What remains unverified?

Successful authorized cloud workload execution, present-day availability forecasts, Azure/GCP VM provisioning and public-service security. The AWS trial was blocked by organization policy.

Help improve the product

Is the historical-versus-live distinction clear?
Which workload constraints and comparisons would you need?
Which labels, layout or evidence are confusing?
What would you require before trusting a cloud recommendation?

Send your feedback directly to the person who shared this demo. This page collects no feedback or personal data.