pamone-obs (Python)

Reusable helper for the pamone obs-agent sidecar. Apps install this once and get logs, resource metrics and (optionally) traces shipped to pamone — with no hostPath (works under Pod Security restricted) and no change to Datadog.

• logs — mirror stdout/stderr into the shared file (fixes stdout-only apps).
• metrics — read the app's OWN cgroup v2 → CPU/memory/IO + /proc/net/dev.
• traces — mirror spans from the tracer already in use (dd-trace or OTel), additively; the Datadog path is untouched.

Install

pip install pamone-obs              # metrics + logs (+ trace auto-detect)
# optional extras carry the tracer libs if the app does not already have them:
pip install "pamone-obs[trace]"     # ddtrace
pip install "pamone-obs[otel]"      # opentelemetry-sdk

Use

Explicit (one line), after the app's logger/tracer setup:

import pamone_obs
pamone_obs.init()

Zero-code (recommended for images): install once, then it auto-starts:

pamone-obs install-pth              # writes zz_pamone_obs.pth into site-packages

Choose features per app (env)

Examples:

• metrics + logs only (no tracer / no DD): PAMONE_TRACES_ENABLED=false
• with traces (pod uses dd-trace): leave traces enabled (or set PAMONE_TRACES_BACKEND=ddtrace).

Sidecar pairing

Set the agent env to the same paths (see the Helm library chart deploy/helm/pamone-sidecar and CONTRACT.md):

AGENT_LOG_PATH=/var/log/pamone/app.log
AGENT_METRICS_PATH=/var/log/pamone/metrics.jsonl
AGENT_SPANS_PATH=/var/log/pamone/spans.jsonl

Build & distribute to clusters

Build a wheel and publish it so pip install pamone-obs works everywhere:

python -m build            # -> dist/pamone_obs-0.1.0-py3-none-any.whl

Options, best → simplest:

1. Internal PyPI (devpi / pypiserver / Nexus) reachable by cluster build runners: twine upload the wheel once; apps do pip install pamone-obs.
2. Bake into an internal base Python image (e.g. playcourt/python fork): install the wheel + pamone-obs install-pth once; every app inherits it and only needs the env + shared volume → truly zero-code.
3. Vendor the wheel into each app image when there is no PyPI egress: COPY pamone_obs-0.1.0-py3-none-any.whl . && pip install ./pamone_obs-*.whl.

The library is stdlib-only; tracer packages are optional extras.

Notes / limitations

• The log tee captures the current process only (child processes have their own streams). For multi-worker servers, each worker runs init() — metrics are duplicated (harmless) and each worker's output is written.
• Calls are best-effort and never raise; failures are logged at debug level.