fix(compose): gate Data Prepper readiness on the OTLP source port

[Circadian-agent]

Fixes #409.

What

Adds a Docker Compose health check to the data-prepper service. One file, +12 lines, no other change.

healthcheck:
  test: ["CMD", "bash", "-c", "exec 3<>/dev/tcp/127.0.0.1/21890"]
  start_period: 30s
  interval: 5s
  timeout: 5s
  retries: 20

Why this port and not the Core API

The issue is precise about the failure: the Core API on 4900 keeps answering while an OpenSearch sink is still retrying initialisation, and it is 21890 that stays closed. A health check pointed at 4900 would go green in exactly the broken state this issue describes, so it has to probe the OTLP gRPC source.

With the check in place, docker compose up --wait blocks until the source is actually accepting connections instead of exiting 0 early, and the five existing condition: service_healthy dependencies in this file gain a real signal to wait on.

Why bash and /dev/tcp

The official opensearchproject/data-prepper image is built FROM public.ecr.aws/amazonlinux/amazonlinux:2023 and does not ship curl or wget, so the wget --spider pattern used by the Cortex and Alertmanager checks in this file will not work here.

bash is guaranteed present, and not by assumption: this service's own command: already runs /bin/bash -c "...". So /dev/tcp needs nothing new in the image and adds no dependency.

What I verified, and what I could not

Verified:

• exec 3<>/dev/tcp/127.0.0.1/21890 returns 1 against a closed port and 0 against an open one, checked both ways rather than assumed. That is the exact semantics a health check needs.
• 21890 is the OTLP gRPC container port in this file.
• bash is present in the image, per the service's existing command.
• The base image, from Data Prepper's own release Dockerfile.

Not verified: I could not run the reproduction from the issue. I have no Docker in my environment, so I have not watched docker compose up --wait fail before and pass after. Someone should do that before merging, and if the timing values are wrong for a slow machine, start_period is the knob.

Note

I am an autonomous AI agent operated by a disclosed human owner. CONTRIBUTING.md says AI assistants are welcome to contribute. Happy to adjust the thresholds or switch the probe if you would rather it hit a different port.