Rust 内核 · Apache-2.0 · 生产可用 Rust core · Apache-2.0 · Production ready

把可观测性
推到离用户最近的地方 Push observability
to where users are

VYQON 把 trace、metric、log 收进同一个内核，编译成单个二进制。冷启动 3ms，常驻内存 18MB，在任何边缘节点上都能跑。 VYQON unifies traces, metrics and logs in one Rust core, shipped as a single binary. 3ms cold start, 18MB resident — it runs on any edge node.

curl -fsSL https://vyqon.me/install.sh | sh

5 分钟接入Get started 看架构Read the design

3ms: 冷启动; cold start
18MB: 常驻内存; RSS
1.2M/s: 单核吞吐; per core
0-dep: 运行时依赖; runtime deps

// 01

为什么再造一个引擎Why another engine

传统可观测性栈把采集、传输、存储拆成三个进程、四种协议、五份配置。你在中心机房调好的一切，搬到边缘就会因为内存和冷启动而全线崩塌。VYQON 的出发点是：边缘不是数据中心的缩小版。 Classic stacks split collection, transport and storage into three processes, four protocols and five config files. Everything tuned for a central datacenter collapses at the edge on memory and cold start. VYQON starts from one premise: the edge is not a small datacenter.

所以我们把三个职责压进一个内核，用 Rust 写成零依赖二进制：没有 sidecar，没有 agent 编排，没有“先把数据发到中心再想怎么存”。采样、聚合、降维在数据产生的那台机器上就完成。 So we compressed all three into one core and wrote it as a dependency-free Rust binary: no sidecar, no agent orchestration, no “ship everything to the centre and figure out storage later”. Sampling, aggregation and reduction happen on the machine where the data is born.

单二进制，无 libc 之外依赖Single binary, no deps beyond libc
OTLP / Prometheus / 自研二进制协议全支持OTLP, Prometheus and a wire-efficient binary protocol
断网时本地 WAL 缓冲，恢复后自动补传Local WAL buffering, automatic replay after a partition
尾部采样在本地完成，省 90% 带宽Tail sampling locally — up to 90% less egress
WASM 插件沙箱，扩展无需重编译WASM plugin sandbox, extend without recompiling

// 02

能力Capabilities

Unified Ingest

同一管线吃下 OTLP、Prometheus 抓取与结构化日志，字段自动归一，不用维护三套采集器。One pipeline for OTLP, Prometheus scrapes and structured logs. Fields are normalised automatically — no three collectors to babysit.

Local Sampling

基于规则的头部/尾部采样在本地执行，错误链路 100% 保留，正常链路按预算砍。Rule-driven head and tail sampling runs locally: 100% of error traces kept, healthy traffic trimmed to budget.

Durable Buffer

本地 WAL + 背压感知的重放器。出口断 6 小时，数据一条不丢，也不会把磁盘写爆。A local WAL with a back-pressure aware replayer. Six hours of egress outage loses nothing and never fills the disk.

Query & Alert

内建 PromQL 子集与流式告警规则，规则热加载，不需要额外的 Alertmanager。A PromQL subset plus streaming alert rules, hot-reloaded — no separate Alertmanager.

WASM Plugins

用任意能编译到 WASM 的语言写处理器，沙箱隔离、毫秒级加载，升级不用重启进程。Write processors in any language that targets WASM. Sandboxed, millisecond loads, no process restart to upgrade.

Secure by Default

mTLS 双向认证、按租户的密钥隔离、审计日志。CDN 级默认安全，不用额外加层。Mutual TLS, per-tenant key isolation, audit logging. Secure defaults out of the box rather than an extra layer to bolt on.

// 03

架构Architecture

数据从产生到落地只经过一次内存拷贝。中间任何一段掉线，都不会让上游阻塞。 From producer to storage the data is copied once. Any segment going down never blocks upstream.

01Producers应用 · 网关 · IoTapps · gateways · IoT
→
02VYQON Core解析 · 归一 · 采样parse · normalise · sample
→
03WAL本地持久缓冲durable local buffer
→
04Shipper压缩 · 批量 · 重放compress · batch · replay
→
05Backends对象存储 / 时序库object store / TSDB

vyqon.toml

[[pipeline]]
name   = "edge"
listen = "0.0.0.0:4317"
protocols = ["otlp", "prometheus"]

[pipeline.sample]
mode    = "tail"
keep_on = ["status=error", "duration>500ms"]
budget  = 0.15

[buffer]
wal_dir = "/var/lib/vyqon/wal"
max_size = "2GiB"

[[exporter]]
type = "s3"
bucket = "vyqon-cold"
compress = "zstd"

plugin.rs

#[vyqon::plugin]
fn on_span(mut s: Span) -> Action {
    // 给每条 trace 打上发布标记
    s.tag("release", env!("GIT_SHA"));

    // 内部健康检查不进存储
    if s.name.starts_with("healthz") {
        return Action::Drop;
    }
    Action::Keep(s)
}

// 04

基准Benchmarks

4 vCPU / 8GB 边缘节点，持续注入 1.2M spans/s，对比同机部署的主流采集器。数字越小越好。 4 vCPU / 8GB edge node, 1.2M spans/s sustained, compared against mainstream collectors on identical hardware. Lower is better.

VYQON

18 MB

Collector A

94 MB

Collector B

140 MB

Agent C

210 MB

复现脚本与原始数据在仓库 bench/ 目录，MIT 许可，欢迎自行复测。 Reproduction scripts and raw data live in bench/, MIT licensed. Re-run it yourself.

// 05

快速开始Quickstart

curl -fsSL https://vyqon.me/install.sh | sh
vyqon init --config vyqon.toml
vyqon serve

# 另开一个终端看看数据进来了没
vyqon top

docker run -d --name vyqon \
  -p 4317:4317 -p 4318:4318 \
  -v $PWD/vyqon.toml:/etc/vyqon/vyqon.toml:ro \
  -v vyqon-wal:/var/lib/vyqon/wal \
  ghcr.io/vyqon/vyqon:0.9.4

helm repo add vyqon https://charts.vyqon.me
helm upgrade --install vyqon vyqon/vyqon \
  --namespace observability --create-namespace \
  --set daemonset.enabled=true \
  --set buffer.maxSize=2Gi

// 06

路线图Roadmap

0.7Rust 内核重写，体积降到 1/6Rust core rewrite, 6× smaller
0.8尾部采样与 WAL 重放Tail sampling and WAL replay
0.9WASM 插件沙箱（当前）WASM plugin sandbox (current)
1.0稳定 API 与 LTS 分支Frozen API and an LTS branch
1.2分布式查询下推Distributed query pushdown

在你的边缘节点上跑起来Run it on your edge node