xtraceAI Observability Service

Collect, store, and query traces, spans, and metrics across LLM and agent workflows. Built in Rust for speed and reliability.

Get Started →

Xinference Integration →

View on GitHub

🔭

Traces & Spans

Full request-chain visibility with nested span trees. Debug P3 cycles (Plan → Execute → Reflect) and multi-step agent workflows.

📊

Time-Series Metrics

Ingest and query GPU utilization, KV cache, token usage, and custom metrics with label-based filtering and downsampling.

⚡

Built in Rust

Axum + Tokio async runtime, batched ingestion via mpsc channels, PostgreSQL storage. Sub-millisecond overhead on the write path.

🔗

OpenTelemetry Compatible

OTLP/HTTP ingestion endpoint with JSON and Protobuf support. Drop-in replacement for Langfuse-compatible instrumentation.

🏷️

Multi-Dimensional Labels

Attach session_id, model_name, agent_role, and arbitrary labels to metrics. Filter and group by any dimension.

📈

Percentile Analytics

Query p50/p90/p99 latency distributions. Compare performance across versions, models, and agent configurations.

What's New

v0.1.0Apr 2026

Xinference Compatibility Hardening — Added Python compat and integration guidance for replacing Langfuse with xtrace, deployment examples for GCP and systemd, smoke / end-to-end validation scripts, and improved batch ingest compatibility for sessionId-style clients.

Integration Guide →SDK Docs →

v0.0.13Feb 2026

Percentile Aggregation & group_by — Metrics query now supports agg=p50|p90|p99 via PostgreSQL percentile_cont, and group_by to split time-series by any label key. Frontend dashboard with trace list and detail viewer.

Metrics API →

v0.0.11Jan 2026

Per-Token Rate Limiting — Per-bearer-token query rate limiting using the governor crate (token bucket). Configurable via RATE_LIMIT_QPS and RATE_LIMIT_BURST. Returns 429 Too Many Requests with a Retry-After header.

v0.0.8Jan 2026

Nebula Signal Contract — Formalized metrics signal contract for Nebula (GPU cluster) integration. Node reports GPU/KV-cache/queue metrics; Router reads real-time signals for load-aware routing; Scheduler reads for placement decisions.

Nebula Docs →