Architecture

How Rezolus is structured and how the pieces fit together.

What you can do with Rezolus: the rezolus agent collects from your systems (CPU, GPU, kernel, containers, services). From there you can expose Prometheus metrics with rezolus exporter, watch a live dashboard with rezolus view, or capture a recording with rezolus record or rezolus hindsight. Recordings are .dendro archives (older ones .rez archives or .parquet files) you can explore in the viewer, analyze with an AI assistant via rezolus mcp, or manage with rezolus recording.
Solid boxes are all the same rezolus binary — you just pick the subcommand for the job. Dashed boxes are what you point it at: the systems and services being measured, which are not part of Rezolus.

Operating Modes

Rezolus is a single binary with six subcommands. Each mode handles a different stage of the observability workflow.

monitoring

Agent

sudo rezolus <config.toml>

The core collection engine. Attaches eBPF programs to kernel tracepoints and kprobes to collect metrics across CPU, scheduler, block I/O, network, memory, TCP, and syscalls. Exposes data via HTTP on port 4241 in JSON and msgpack formats.

upload

Exporter

rezolus exporter <config.toml>

Pulls metrics from the agent's msgpack endpoint and serves them in Prometheus format on port 4242. Converts histograms to configurable summary percentiles. Designed for Prometheus scraping.

fiber_smart_record

Recorder

rezolus record [-o out.dendro]

On-demand metric capture to a .dendro archive, the only format it writes. Connects to a running agent and writes high-fidelity snapshots at configurable intervals and durations. Ideal for benchmarking and performance engineering.

history

Hindsight

rezolus hindsight <config.toml>

Maintains a rolling buffer of high-resolution metrics on disk as a .dendro archive. Trigger snapshots via SIGHUP or the HTTP API. Supports time-range queries for extracting the window around an incident.

dashboard

Viewer

rezolus view <input> [listen]

Web-based interactive dashboard for exploring .dendro, .rez and Parquet recordings or streaming live from a running agent. Includes a built-in TSDB with PromQL query support and pre-built dashboards for each metric category.

smart_toy

MCP Server

rezolus mcp [subcommand]

AI-guided analysis server. Provides tools for querying recordings with PromQL, detecting anomalies, analyzing correlations between metrics, and describing recording metadata, plus tools that write findings back as timeline events. More.

eBPF Integration

Rezolus uses libbpf-rs for eBPF program management. BPF programs are written in C (.bpf.c files) alongside their Rust module counterparts and compiled at build time via build.rs.

At runtime, the agent attaches these programs to kernel tracepoints and kprobes. This provides kernel-level visibility with near-zero overhead — no polling of /proc or /sys required for instrumented metrics.

Requires: Linux kernel 5.8+ and root access. macOS builds omit eBPF samplers.

Key Technical Details

  • • Async runtime: Tokio with minimal threads
  • • HTTP framework: Axum
  • • Metrics library: metriken crate
  • • Sampler registration: linkme distributed slices for static registration
  • • Data format: .dendro archives via the dendro crate
  • • Hardware topology: systeminfo crate
  • • Architectures: x86_64 and ARM64
← MCP Server Metrics →