-
Notifications
You must be signed in to change notification settings - Fork 9
Support
How to get a problem looked at, and how to do it without putting anything sensitive in a public place.
| Your situation | Where to go |
|---|---|
| An error you can describe without your own data | Open a GitHub issue |
| Anything involving your logs, queries, documents, hostnames or topology | Your Elastic support channel, not GitHub |
| A commercial license for a CC-BY-NC-4.0 model, or renewal | Elastic sales |
| A question about how long a model is supported | Product & Model Lifecycle |
The GitHub tracker is public and permanent. A container log line can contain a request, a document, an internal hostname or a model id you would rather not publish, and an issue cannot be un-published. If there is any doubt, use your support channel instead — support for a licensed deployment tracks your Elastic support entitlement and is delivered through your approved channel, which never requires the deployment to reach the internet.
Check Troubleshooting first. It covers the errors that account for most reports, including several whose fix is not obvious from the message.
The more of this you can share, the less back-and-forth:
# Which image, exactly
docker inspect --format '{{index .RepoDigests 0}}' <image>
# What the server thinks it is
curl -s http://localhost:8080/health
# The host
docker version --format '{{.Server.Version}}'
nvidia-smi --query-gpu=name,driver_version,memory.total --format=csv # GPU onlyPlus:
- What you called and what came back — the endpoint, the HTTP status, and the shape of the request. Not the contents, unless they are safe to share.
- Whether it is reproducible, and whether it happens on a smaller input.
- What changed — a new image, a new host, a new model, a different request shape.
-
Container logs if they are safe to share.
docker logs <container>is the standard ask, so see below for what is in them.
Worth knowing before you paste them anywhere:
- The server's own log line per request is counts and timings only — how many inputs or documents, how many tokens, elapsed milliseconds, throughput. No request content.
- The access log records method, path and status, so it shows which endpoint was called but not what was sent.
- Startup logs include the model id, device, dtype, thread count and the license key state.
- A traceback from an unexpected error is the exception: it can include values from the failing call. That is the part to read before sharing.
A safe minimum for a public issue is usually the last few lines around the error plus the /health output, with anything identifying replaced.
Do not open a public issue for a suspected vulnerability. Use your Elastic support channel, or Elastic's security contact, and include the image digest and the smallest reproduction you have.
For questions about what the container does on a network — rather than a suspected flaw — Security & Hardening is written to be quoted directly into a questionnaire.
- Troubleshooting - the errors that come up most, with fixes
- Security & Hardening - network behaviour, authentication, and the hardening checklist
- FAQ - licensing, support horizon, and other non-technical questions
Repo - Issues - Releases - Commercial license