跪拜 Guibai
← All articles
Backend

SkyWalking Ships a Three-Layer AI Stack That Reads Traces So You Don't Have To

By 苏三说技术 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Distributed tracing has always produced more data than most on-call engineers can interpret under pressure. SkyWalking's AI assistant collapses the 20-minute drill of flipping through trace lists, reading span trees, and correlating logs into a single question, while the GenAI dashboard gives teams that ship LLM-powered features their first real view of per-call cost and latency.

Summary

SkyWalking's new AI capabilities ship across three independent layers. Horizon UI AI Assistant answers natural-language questions by querying live OAP data and rendering real charts, not just text summaries. Virtual GenAI instruments every LLM call made through Spring AI or OpenAI SDKs, surfacing latency percentiles, token counts, TTFT, and estimated cost per model. AI Pipeline connects to an external gRPC service for ML-based baseline calculation and URI pattern recognition, flagging anomalies without manual thresholds. Each layer can be adopted separately, and the assistant defaults to off so teams control which model endpoint it hits — local Ollama or a cloud API — keeping data in-region.

Takeaways
Horizon UI AI Assistant answers natural-language questions by pulling live OAP data and rendering real charts inline, not text-only replies.
Virtual GenAI captures latency, token usage, TTFT, and estimated cost for every LLM call made via Spring AI or OpenAI SDK.
AI Pipeline uses an external gRPC AI service to compute adaptive baselines and recognize RESTful URI patterns, triggering alerts without manual thresholds.
All three AI layers are independent; teams can start with the assistant without touching GenAI or Pipeline.
The assistant is off by default and can point to a local Ollama instance, keeping telemetry data from leaving the network.
GenAI cost estimation requires a manual pricing config in gen-ai-config.yml per provider and model.
Virtual GenAI needs Java Agent >= 9.7; AI Pipeline needs a separate deployed service implementing the gRPC protocol.
Conclusions

SkyWalking's AI strategy is additive, not a rewrite — each layer bolts onto the existing OAP and agent architecture, which lowers the adoption risk for teams already running it in production.

Making the assistant default-off and model-agnostic sidesteps the enterprise objection that AI features will leak observability data to a third-party cloud.

GenAI monitoring addresses a genuine blind spot: most teams deploying LLM features have no per-request cost or latency visibility, so they optimize prompts and model selection by guesswork.

The assistant's 'show, don't describe' design — rendering real charts instead of hallucinated text — is a practical hedge against LLM confabulation in an operational context where wrong numbers trigger wrong decisions.

Concepts & terms
Horizon UI
SkyWalking's next-generation web console, released as a standalone container image. It replaces the classic UI and hosts the AI Assistant.
OAP (Observability Analysis Platform)
SkyWalking's backend service that receives, analyzes, and aggregates telemetry data from agents. The AI Assistant queries it through the same protocol as dashboards.
Virtual GenAI
A SkyWalking 10.4 capability that instruments LLM calls from Spring AI or OpenAI SDK, reporting traffic, latency, token usage, TTFT, and estimated cost as trace-derived metrics.
AI Pipeline
A SkyWalking module that communicates with an external AI service over gRPC to perform ML tasks like baseline calculation and HTTP URI pattern recognition on telemetry data.
TTFT (Time to First Token)
The latency between sending a request to an LLM and receiving the first token of the response. A critical UX metric for streaming chat applications.
R3 Algorithm (RESTful Pattern Recognition)
A machine-learning algorithm used by SkyWalking's AI Pipeline to automatically group and classify RESTful API URL patterns, e.g., collapsing /users/123 and /users/456 into a single pattern.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗