Elina Python SDK

Python tracing SDK for self-hosted Elina. Licensed under MIT.

Install

python -m pip install elinaobs

For development from this repository:

python -m pip install -e '.[dev]'
pytest -q

Public API

from elinaobs import init, trace, flush, shutdown, diagnose, submit_feedback

init(api_key="...", endpoint="http://localhost:3000/v1/traces", service_name="my-agent")

@trace(name="Agent", as_type="agent")
async def run(question):
    return await your_code(question)

# The existing OTel-instrumented operations called inside run() are children.

Application feedback is optional. After a trace has finished, an app with a verdict can send it using the same project key. Inside a decorated function, capture the trace ID and keep it with the result:

from opentelemetry import trace as otel_trace
trace_id = f"{otel_trace.get_current_span().get_span_context().trace_id:032x}"
# Later, after the traced function returns:
submit_feedback(trace_id, "bad", note="Wrong answer")  # trace_id is 32 lowercase hex characters

Apps without a feedback signal do not call this function. Feedback is stored separately from the dashboard's human review. Keep the project key in server-side code, not in a browser.

To send a demo trace to the local Elina server, create a project in the dashboard and copy its one-time ingestion key. Then, from this directory in PowerShell:

$env:ELINA_WRITE_KEY = "paste-project-ingestion-key"
$env:ELINA_BASE_URL = "http://127.0.0.1:3000"
python examples/agent.py

The script prints the trace ID and reports whether export succeeded. Open the project’s Traces page and select that ID to inspect the three spans and their captured input/output.

Only operations that are explicitly decorated or already instrumented by a compatible OpenTelemetry instrumentation package create spans. No framework integrations are installed, imported, or configured by Elina.

The Elina receiver lives in the iamrobins/elina repository. Start it before running examples/agent.py. The default tests use in-memory and local HTTP receivers.

To also verify export through a real OpenTelemetry Collector, set OTELCOL_BINARY to the Collector executable path and run python -m pytest -q tests/test_collector.py. This optional test starts a local Collector with a file exporter and checks the received spans.

What it does

• Sync + async function decorator with native OTel parent context.
• 10 span display classifications using elinaobs.observation.type.
• Selective export: only spans starting inside a @trace() scope.
• OTLP/HTTP protobuf, bearer-token auth, background batching.
• Optional bounded input/output capture, disabled by default.
• An optional redact callback can remove secrets from captured free text before export.
• attribute_allowlist and attribute_denylist filter span, event, and link attributes in Elina exports; the denylist wins when both contain a key.
• Exceptions automatically record OTel error status/events.
• Shares an existing OpenTelemetry SDK TracerProvider when present.
• flush() reports failed export batches or dropped spans. diagnose() returns failed_batches, dropped_spans, and retry_attempts without exposing credentials. Drop counts are updated after a successful processor flush; retry attempts count extra HTTP requests made for a batch.

Non-goals for v0.1

• Framework adapters or integrations=[...].
• Raw threads, multiprocessing, distributed orchestration, or generator tracing.
• Automatic patching of uninstrumented Python libraries.
• Prompt management, evaluations, or server/UI implementation.

Known limitations before production

• SDK processor cannot modify existing OTel provider sampling or resources.
• Capture filtering uses a ContextVar marker; parent spans begun before a decorator scope are not exported to Elina unless also eligible.
• as_type="event" is a timed span category, not an OTel instant event.
• Incoming external spans from opentelemetry-instrument share global provider and may require controlling initialization order.
• Bounded batches can drop data on overload. dropped_spans is current only after flush() or shutdown(). retry_attempts is None when an OpenTelemetry credential-provider session is configured. Export failures remain counted even if later batches succeed.
• Input/output redaction is best-effort only. Explicitly inspect anything you opt in to collect.
• Attribute lists do not change an application's other OpenTelemetry exporters or resource metadata. Keys are matched exactly.
• shutdown() ends only Elina's processor. Do not reinitialize the SDK in the same interpreter after shutdown.

See docs/LLD.md for the finalized architecture and build checklist.