Clarvynn OpenTelemetry Adapters

Source Governance for OpenTelemetry — policy-based control over what telemetry gets exported.

Status: Early Stage

Clarvynn is an early-stage project exploring Source Governance for OpenTelemetry. The architecture is validated, but the implementation is evolving. We welcome feedback, testing, and contributions to help mature the codebase.

---

What is Clarvynn?

Clarvynn provides Deferred Head Sampling for OpenTelemetry - sampling decisions made after request completion, so you never miss critical signals.

Define your important telemetry (errors, slow requests, key business paths) in a simple YAML policy. Clarvynn ensures these are always captured while applying configurable sampling to routine traffic.

The goal: High-fidelity signal with lower volume. Early testing shows potential for 60-80% reduction in telemetry volume while preserving visibility into the requests that matter.

---

Language Support

Language	Status	Adapter
Python	Available	adapters/opentelemetry-python/
Java	Planned	-
Node.js	Planned	-
Go	Planned	-

---

Quick Start (Python)

1. Install

pip install clarvynn

2. Create Policy

# policy.yaml
sampling:
  base_rate: 0.01  # 1% of routine traffic

conditions:
  - name: errors
    when: "status_code >= 400"
  - name: slow_requests
    when: "duration_ms > 1000"

3. Run

export CLARVYNN_ENABLED=true
export CLARVYNN_POLICY_PATH=policy.yaml
opentelemetry-instrument python app.py

That's it. Clarvynn is now intelligently sampling your telemetry.

---

Key Features

• Deferred Head Sampling - Decisions made after request completion
• Policy-driven - Simple YAML configuration
• Minimal overhead - < 50µs per span evaluation
• Drop-in - Works with existing OTEL instrumentation
• Distributed tracing - W3C TraceContext Level 2 support
• Signal Improvement - High-Fidelity Signal, Low-Volume Noise
• Log correlation - Logs automatically follow trace sampling

---

How It Works

Traditional sampling makes decisions before execution:

HEAD-BASED:  [Sample?] → Execute → [Don't know if it failed!]
                - Might drop errors

Clarvynn uses Deferred Head Sampling to decide after execution:

DEFERRED HEAD:  Execute → [Error? Slow?] → [Sample!]
                - Never miss critical signals

Example:

# Your policy
conditions:
  - name: errors
    when: "status_code >= 500"
  
# What happens:
Request → Execute → Status: 500 → Exported (error captured)
Request → Execute → Status: 200 → 1% chance (base_rate)

---

Documentation

Getting Started:

• Quick Start - 5-minute tutorial
• Python Adapter - Python-specific docs

Python Adapter Docs:

• Configuration - Python config reference
• Frameworks - Flask, Django, FastAPI
• How It Works - Python architecture

Shared Documentation:

• CPL Reference - Policy language (all languages)
• W3C TraceContext Level 2 - Distributed tracing
• Architecture - System design
• Adapter Development - Building new adapters

---

Architecture

clarvynn/
├── core/                          # Shared CPL engine (all languages)
│   ├── cpl_engine/                # Policy evaluation logic
│   ├── policies/                  # Example policies
│   └── specs/                     # Language-agnostic specs
│
└── adapters/                      # Language-specific adapters
    ├── opentelemetry-python/      # Available
    ├── opentelemetry-java/        # Planned
    ├── opentelemetry-nodejs/      # Planned
    └── opentelemetry-go/          # Planned

Design Philosophy:

• core/ = Language-agnostic policy engine and specifications
• adapters/ = Language-specific OpenTelemetry integrations

Each adapter uses the same CPL policy format, ensuring consistent behavior across languages.

Non-Blocking Architecture (Python)

Clarvynn operates as a Lightweight Adapter that ensures minimal impact on application performance by using a fully non-blocking export pipeline:

Request Ends
  ↓
ClarvynnSpanProcessor (Sync, ~2µs)
  ├─ 1. Evaluate CPL Policy (Keep/Drop)
  └─ 2. Inject TraceState (W3C Level 2)
  ↓
BatchExporterAdapter (Async Queue)
  ↓
Background Thread (Async)
  ↓
OTLPSpanExporter (Network I/O)

This architecture ensures that governance decisions happen synchronously (to decide before queuing), but network I/O happens asynchronously (to avoid blocking the application).

---

Example: Multi-Service Deployment

Deploy the same policy across all services:

# policy.yaml (use everywhere)
sampling:
  base_rate: 0.01

  conditions:
    - name: errors
      when: "status_code >= 400"
    - name: slow_requests
    when: "duration_ms > 1000"

What happens:

API Gateway (Python):
  └─ Error (500) → Exported + marked critical

Auth Service (Java):       # Coming Soon
  └─ Sees critical trace → Exported

Database Service (Go):     # Coming Soon
  └─ Sees critical trace → Exported

Result: Complete trace captured end-to-end

---

Testing

Clarvynn has comprehensive test coverage (341 tests, 100% pass rate):

# Install test dependencies
pip install -r requirements-dev.txt

# Run all tests
pytest

# Run with coverage
pytest --cov=core --cov=clarvynn --cov-report=html

See tests/README.md for detailed testing guide.

---

Compatibility

Component	Supported Versions
Python	3.9, 3.10, 3.11, 3.12, 3.13
OpenTelemetry SDK	1.25.0 - 1.39.0 (tested)
OpenTelemetry API	1.25.0 - 1.39.0

Compatibility is verified with automated testing. See scripts/test_otel_compatibility.py.

---

Contributing

We welcome contributions! See CONTRIBUTING.md.

Adding a new language adapter? See Adapter Development Guide.

---

License

Apache 2.0 - See LICENSE