Docker execution model

HyperOptimizer runs your workload as containers. The image owns the code path; the platform owns scheduling, parameter selection, and metric collection.

Container contract

Your image must be able to run one trial from its default command (or the command configured in the experiment). HyperOptimizer appends or substitutes managed HPO arguments.

python main.py --hpo-lookback-window=50 --hpo-risk-multiplier=1.4

The process should exit after the trial finishes. A successful run exits with code 0 and prints metric lines.

Images source

On Free, Starter, and Pro, the image must come from your org Images registry. Publish with Publish an image. Enterprise can allow custom Docker images outside the org registry.

Parallel execution

Multiple trials can run at the same time (subject to plan parallelism). Treat each container as isolated. If trials write files, use unique output paths (for example include HYPEROPTIMIZER_TRIAL_ID) or write only within the container filesystem.

Note

Parallelism speeds up experiments, but the optimizer can only learn from completed trials. Emit metrics as soon as the trial result is known.

Inputs and outputs

Inputs

• Docker image (from Images, unless Enterprise custom Docker)
• Managed CLI parameters (--hpo-* or plain --*)
• Platform env: HYPEROPTIMIZER_TRIAL_ID, HYPEROPTIMIZER_EXPERIMENT_ID, HYPEROPTIMIZER_ORG_ID
• Optional custom environment and data access baked into the image

Outputs

• Exit code
• Stdout metric lines
• Logs for debugging failed trials

Build the image with the Docker image guide.