How Alibaba Cloud Function Compute Uses OpenTelemetry for Full‑Stack Tracing

Background

Request‑level tracing is a core capability for diagnosing performance bottlenecks in distributed systems. Alibaba Cloud Function Compute (FC) has replaced its previous Jaeger 2.0 implementation with the OpenTelemetry W3C standard. This upgrade provides full‑path observability of function execution, turning the traditional serverless “black‑box” into a transparent, traceable component.

Key Features

FC system‑level span propagation : Critical lifecycle events of internal components (scheduler, cold‑start module, etc.) are emitted as OpenTelemetry spans, covering the entire function lifecycle from dispatch, initialization, execution to release.

Automatic stitching of business spans : User‑defined spans created inside function code are automatically linked to the FC system spans, producing an end‑to‑end trace that highlights hotspots such as cold‑start latency or resource contention.

Standard W3C header support : The implementation respects traceparent, tracestate and baggage headers, ensuring lossless context propagation to downstream services (databases, message queues, etc.).

Cross‑environment interoperability : Trace context can be passed across functions, services and even cloud providers, allowing seamless integration with existing OpenTelemetry tooling.

Dynamic cost control via sampling : Sampling rates are configurable per function (e.g., 1 % for routine monitoring, 100 % for incident investigation) to balance data volume against resource overhead.

Usage Scenarios

After enabling tracing, developers can view function execution details in both the FC console and any OpenTelemetry‑compatible tracing UI. A typical example is a LangChain client invoking a Gaode weather service; the trace shows cold‑start time, SSE connections, message streams and agent calls.

Diagnostic steps :

Inspect the PrepareCode span to evaluate cold‑start duration; if it exceeds expectations, reduce the deployment package size.

When using custom runtimes or container images, monitor the RuntimeInitialization span; unusually long initialization indicates the need to optimise startup scripts.

AI‑Assisted Debugging

If an abnormal request is detected, the corresponding trace can be examined for error details. The AI Operations Assistant can further analyse the trace to pinpoint root causes, accelerating incident resolution.

LLM Application Monitoring

Installing the ARMS Python probe on Large Language Model (LLM) applications enables call‑chain analysis. The tracing UI displays spans for input processing, output generation, token consumption and other metrics, providing fine‑grained visibility into LLM workloads.

Effect Comparison

Before the upgrade , tracing relied on log aggregation, making it impossible to separate system latency from business latency and causing context loss across services.

After the upgrade , spans are visualised with clear segmentation, allowing precise bottleneck identification. W3C header propagation restores complete end‑to‑end trace continuity across heterogeneous environments.

Summary

The deep integration of Function Compute with OpenTelemetry delivers full‑stack trace transparency, covering both system‑level and business‑level spans. By adhering to the unified W3C protocol, data silos are eliminated and cross‑environment consistency is ensured. Dynamic sampling provides an economical observability solution, while AI‑assisted diagnostics and the ARMS Python probe extend troubleshooting capabilities for complex workloads such as LLM applications.