Troubleshooting

Use this page when a trial starts but does not produce the result you expected — or when new trials never start.

No metrics collected

Check stdout for lines that match the exact metric format.

hpo.metrics.<key>=<json>

Common causes:

• The metric prefix is missing or misspelled.
• The value after = is not valid JSON.
• The program wrote metrics to a file but did not print them.
• The process crashed before metric emission.
• Objective names in the dashboard do not match the keys you printed.

Unknown argument

The dashboard parameter slug must match your parser.

Name mapping

Keep these three forms aligned for every parameter (default --hpo-* mode):

Dashboard: lookback-window
CLI:       --hpo-lookback-window=50
Python:    args.hpo_lookback_window

In plain flag mode, the CLI form is --lookback-window=50 instead.

Image pull or registry errors

• Confirm the image exists under Dashboard → Images for your org project.
• Re-copy login credentials and push again if the tag is missing.
• Free/Starter/Pro cannot use arbitrary public images; publish to Images first.
• Repeated platform start failures (for example bad ImagePull) can pause suggestion until the image is fixed.

Trial exits non-zero

Read the container logs around the failure. For local reproduction, run the image with representative arguments:

docker run --rm your-image:latest \
  --hpo-lookback-window=50 \
  --hpo-risk-multiplier=1.4

New trials are not starting

Check:

• Credit balance — top up on Plans and credits
• Experiment paused or stopped — Control an experiment
• Failure budget exhausted (max failed trials)
• Experiment spend limit reached
• Plan parallelism saturated

Results look wrong

If metrics are collected but rankings look odd, check:

Objective direction

Confirm whether the experiment should maximize or minimize the metric.

Metric scale

Confirm that your metric units are consistent across trials.

Too few observations

Trading backtests with very few trades can produce misleading scores.

Hidden randomness

Seed random processes or emit the seed as a metric for debugging.