[Bug] Load-balancing endpoints ignore HEALTH_CHECK_PATH (always poll /ping); PORT_HEALTH defaults to 80 instead of PORT

[robster7674]

Summary

On load-balancing Serverless endpoints, two documented settings don't behave as described in Load balancing endpoints:

1. HEALTH_CHECK_PATH is ignored. The load balancer always polls GET /ping. The docs say: "By default the load balancer polls /ping, but you can point it at any path by setting the HEALTH_CHECK_PATH environment variable", and they give llama.cpp's /health as the motivating example.
2. PORT_HEALTH doesn't default to PORT. With PORT=8000 set and PORT_HEALTH unset, the saved template gets PORT_HEALTH=80 injected. The env var table says PORT_HEALTH defaults to "Same as PORT".

Together, these mean the documented use case (stock ghcr.io/ggml-org/llama.cpp:server-cuda, health on /health) never becomes healthy. The worker loads the model and listens, but it's never routed to, and it's killed and restarted after a few minutes.

Reproduction (2026-09-23)

Endpoints created with the GraphQL saveEndpoint mutation, type: "LB", scalerType: "REQUEST_COUNT", workersMin: 0, workersMax: 1, with an inline template. The console "Health check endpoint" field was not used.

1. HEALTH_CHECK_PATH ignored. A request-logging worker:

• imageName: "python:3.12-slim"
• dockerArgs: "python -u -m http.server 80"
• env: [{key: "HEALTH_CHECK_PATH", value: "/health"}]

The platform then injected PORT=80 and PORT_HEALTH=80. The worker log (runpodctl serverless logs <id>) shows only /ping probes, never /health:

100.64.0.76 - - [23/Sep/2026 08:28:53] "GET /ping HTTP/1.1" 404 -
100.64.0.76 - - [23/Sep/2026 08:28:54] "GET /ping HTTP/1.1" 404 -
100.64.0.76 - - [23/Sep/2026 08:28:55] "GET /ping HTTP/1.1" 404 -
...

2. PORT_HEALTH default. The template env was sent as LLAMA_CACHE=/models, PORT=8000, HEALTH_CHECK_PATH=/health. Reading it back with podTemplate(id) returned:

env: LLAMA_CACHE=/models, HEALTH_CHECK_PATH=/health, PORT=8000, PORT_HEALTH=80
ports: 80/http,8000/http

Expected

• The load balancer polls the path in HEALTH_CHECK_PATH when it's set in the template env.
• PORT_HEALTH defaults to the value of PORT.

If either setting is intentionally only honored through the console field or the v2 REST API, the docs should say so.

Impact

Any image whose server doesn't answer /ping (llama.cpp, many OpenAI-compatible servers) needs a custom wrapper image, which is exactly what the docs say HEALTH_CHECK_PATH avoids.