Limits and troubleshooting
Current limits
Section titled “Current limits”| Limit | Value |
|---|---|
| Request body | 1 MiB |
| Workflow YAML | 64 KiB |
| Request rate | 120 requests/minute per client IP, per server process |
| Execution deadline | 30 seconds by default; operator-configurable |
| Run list page | 1–100 results, default 50 |
| Run list offset | 0–10000 |
Health and readiness checks are excluded from rate limiting. Proxy headers are not trusted by default. These are development-service limits, not hosted production guarantees.
Login fails
Section titled “Login fails”Confirm the service is reachable, its Auth0 settings are configured, and your identity has been provisioned into a Banh account. Operators should use the private repository’s AUTH0.md. Retry login after access-token expiration; this revision does not store refresh tokens.
BANH_API_TOKEN overrides saved credentials. Check whether an expired environment token is shadowing a fresh login. An invocation key cannot be used for whoami or deployment.
Workflow or input rejected
Section titled “Workflow or input rejected”Run banh validate workflow.yaml locally. Check unknown fields, decision references, flow expressions, YAML aliases, and the final position of else. For direct deployment, match the URL slug to the YAML process and stay within the source size limit.
Text workflows require a JSON string inside the HTTP request’s input field. CLI --input behavior is documented in workflow syntax.
Every cloud result looks the same
Section titled “Every cloud result looks the same”That is expected with the current fake backend. It returns true for whether, the first option for one_of, and zero for scale. Use local banh run to evaluate real predictions.
Local model setup fails
Section titled “Local model setup fails”The first local run downloads approximately 1.7 GB. Check network access, cache disk space, and roughly 2–3 GB of available RAM. To use an existing complete bundle, pass --model-dir; see CLI options.
Execution times out
Section titled “Execution times out”The service returns HTTP 504 and persists a failed run. In-flight inference cannot currently be cancelled, but late completion does not overwrite the failure. Inspect the run with a developer token.
Runs interrupted by process termination can remain running. Crash recovery is not implemented in this version.
Service is live but not ready
Section titled “Service is live but not ready”/health checks HTTP liveness. /ready additionally checks the database and initial migration. Operators must apply migrations explicitly; startup does not apply them automatically. Development setup and database operations remain in the private cloud repository.