pytracingx 0.2.0

Rust + OpenTelemetry powered Python bindings for traces, metrics, and logs (SLS-first)

pytracingx

中文版 README

Rust + OpenTelemetry powered Python bindings for traces, metrics and logs, with first-class support for Aliyun SLS and ARMS as OTLP backends.

Why pytracingx over Python OpenTelemetry SDK?

	pytracingx (Rust)	opentelemetry-python
Performance	Serialization (protobuf), compression (gzip/lz4), batching, network I/O all happen on Rust native threads — never holds the GIL	Every export step runs on Python threads under the GIL; 5–15% CPU overhead is measurable under heavy span/metric load
Memory	Span/metric data structures live on the Rust heap, zero Python object overhead	Each span is a Python object with dict attributes; significant GC pressure under high traffic
Startup	Single .so file, import pytracingx ~15ms	Pulls in opentelemetry-api + -sdk + -exporter-otlp + grpcio/protobuf and a dozen others, cold start 200–500ms
Dependencies	Zero Python runtime deps (everything compiled into the native module)	Drags in grpcio (C build), protobuf, googleapis-common-protos; wheel size > 50MB
GIL friendliness	start_span() / counter.add() / logger.info() only do FFI argument conversion (μs), then drop the GIL	Python SDK's start_span does context management, attribute serialization and sampler decisions in Python (10–50μs) holding the GIL throughout
Async safety	Span context lives in contextvars, naturally inherited by asyncio.Task; the Rust tracing layer doesn't depend on the Python event loop	BatchSpanProcessor spawns extra daemon threads and uses threading.Event; occasional races when paired with uvloop
Console output	tracing's fmt::Layer unifies span begin/end + log events with compact / pretty / json formats	Needs a separate ConsoleSpanExporter plus a logging handler; two inconsistent formats
Native SLS	Built-in SlsLogSink writes to any logstore directly (protobuf + lz4 + HMAC signing all done in Rust)	Requires aliyun-log-python-sdk or hand-rolled HTTP upload
Single wheel	One abi3-py39 wheel covers Python 3.9–3.13+ across all platforms (manylinux/macOS/Windows)	grpcio wheel must be built per Python version per platform

When to use it

• High-QPS services (>1000 RPS): Rust handles batch export in the background; per-call overhead < 1μs on the Python side
• Latency-sensitive workloads: span export never blocks the GIL
• Containers / serverless: single-file deploy, fast cold start, no grpcio compilation
• Aliyun full stack: traces → ARMS, metrics → ARMS, logs → any SLS logstore — one Config covers all

Observability for the AI era — built for LLM inference

In LLM / large model inference, any CPU-side observability cost translates directly into GPU underutilization. When the Python main thread is busy with trace serialization, log formatting and metric aggregation, GPU kernel launches, KV-cache scheduling and batch composition all get delayed. Symptoms include:

• TTFT (Time To First Token) jitter: span export blocks the GIL and request enqueue is delayed
• GPU bubbles: the CPU can't feed data fast enough, GPU SMs idle, throughput drops
• Degraded batch scheduling: vLLM / SGLang schedulers depend on a low-latency event loop; daemon threads from the Python SDK fight uvloop and break continuous batching

pytracingx's Rust-native design fits these scenarios naturally:

Pain point	Traditional Python SDK	pytracingx
GPU scheduling blocked by GIL	start_span holds GIL for 10–50μs, preempting inference steps	FFI argument copy then GIL released, μs-level return
Token-level tracing overhead	Per-token spans create heavy Python GC pressure	Spans allocated on the Rust heap, zero impact on Python GC
Prompt/completion log volume	Large payloads serialized in Python slow down the main loop	Protobuf + async batching all run inside the Rust tokio runtime
Multi-GPU / multi-process deployment	Each worker loads the full Python OTel stack, doubling memory	Single .so, shared abi3 wheel, < 5MB resident overhead beyond GPU memory

Inference service integration pattern

# At the entry point of vLLM / SGLang / TGI etc.
with ptx.start_span("llm.inference", attributes={
    "llm.model": "qwen2.5-72b",
    "llm.prompt_tokens": prompt_len,
    "gen.batch_size": batch_size,
}) as span:
    output = engine.generate(prompts)        # GPU work runs undisturbed
    span.set_attribute("llm.completion_tokens", output.usage.completion_tokens)
    span.set_attribute("llm.ttft_ms", output.metrics.first_token_ms)

Suggested observation dimensions:

• Trace: end-to-end request → tokenize → schedule → prefill → decode → detokenize
• Metrics: TTFT / TPOT (Time Per Output Token) / GPU SM utilization / KV-cache hit rate
• Logs: sample prompt / completion to SLS for offline RAG / fine-tuning data

Core idea: observability should be part of the AI infrastructure, not a performance tax.

Architecture

Python  ──►  ptx.start_span / ptx.get_logger / ptx.get_meter
                                │
                                ▼
                        tracing crate (Rust)
              ┌──────────────────┴──────────────────┐
              │                                     │
        fmt::Layer                       tracing-opentelemetry
        (terminal)                  + opentelemetry-appender-tracing
                                    + SlsLogLayer (native SLS)
                                                    │
                                                    ▼
                            SdkTracerProvider / SdkLoggerProvider / SdkMeterProvider
                                       opentelemetry-otlp (async reqwest)
                                                    │
                                                    ▼
                                   SLS / ARMS / any OTLP Collector

Each signal is configured via a Sink object. A signal is enabled if its Sink appears in the sinks list.

Installation

pip install pytracingx

Pre-built wheels are available on PyPI for:

• Linux x86_64 / aarch64 — manylinux_2_28 (glibc) and musllinux_1_2 (Alpine)
• macOS — universal2 (Intel + Apple Silicon)
• Windows — x64 (MSVC)
• Python 3.9, 3.10, 3.11, 3.12, 3.13+ (single abi3 wheel)

No Rust toolchain or system OpenSSL is required at install time — everything is statically linked into the wheel.

Quickstart

import asyncio
import pytracingx as ptx

async def main():
    ptx.init(ptx.Config(
        service_name="payment-svc",
        resource_attributes={"deployment.environment": "prod"},
        sinks=[
            # Traces + Metrics → ARMS
            ptx.TraceSink(
                endpoint="http://tracing-xxx.arms.aliyuncs.com/.../api/otlp/traces",
                protocol="http/protobuf",
                sampler="parent_based_traceid_ratio",
                sampler_arg=0.5,
            ),
            ptx.MetricSink(
                endpoint="http://tracing-xxx.arms.aliyuncs.com/.../api/otlp/metrics",
                protocol="http/protobuf",
                export_interval_ms=30_000,
            ),
            # Logs → SLS native (any logstore)
            ptx.SlsLogSink(
                endpoint="cn-hangzhou.log.aliyuncs.com",
                project="my-proj",
                logstore="app-logs",
                ak_id="...",
                ak_secret="...",
            ),
        ],
    ))

    meter = ptx.get_meter("payment")
    logger = ptx.get_logger("payment")
    orders = meter.counter("orders_total")

    with ptx.start_span("checkout", kind="server", attributes={"user.id": "u1"}) as span:
        orders.add(1, attributes={"sku": "abc"})
        logger.info("checkout done", attributes={"sku": "abc"})
        span.set_attribute("amount", 12.34)

    ptx.shutdown()

asyncio.run(main())

Sink Types

Sink	Backend	Protocol	Use Case
TraceSink	Any OTLP collector	gRPC / HTTP	Distributed tracing spans
MetricSink	Any OTLP collector	gRPC / HTTP	Counters, histograms, gauges
OtlpLogSink	Any OTLP collector	gRPC / HTTP	Logs via OTLP (lands in the trace instance's -logs logstore on SLS)
SlsLogSink	Aliyun SLS native API	HTTPS	Logs to any SLS logstore (not limited to the trace instance)

Console-only mode

# No sinks → no network, just terminal output
ptx.init(ptx.Config(service_name="my-app", console_level="debug"))

Context propagation (server side)

# with-syntax auto-restores context on exit
with ptx.extract_headers(dict(request.headers)):
    with ptx.start_span("POST /api/orders", kind="server") as span:
        ...

Examples and benchmarks

• examples/demo.py — comprehensive async example covering traces + metrics + logs + context propagation
• examples/fastapi_middleware.py — FastAPI middleware that creates a server span per request and propagates W3C trace context
• examples/traces_to_xtrace.py — route traces to ARMS while metrics/logs go to SLS
• benchmarks/ — microbenchmarks comparing pytracingx vs opentelemetry-python on the hot path

Bridging stdlib logging

import logging
from pytracingx.logging import SLSLoggingHandler

logging.basicConfig(level=logging.INFO, handlers=[SLSLoggingHandler()])
logging.getLogger("foo").info("hello from stdlib")

Building from source

pip install maturin
maturin develop --release

Requires Rust >= 1.85 (edition 2024) and Python >= 3.9. Wheels are abi3 (abi3-py39).

License

MIT