#tracing #spans #jaeger #datadog #open-telemetry

minitrace

A high-performance timeline tracing library for Rust

11 unstable releases (3 breaking)

0.6.4 Feb 8, 2024
0.6.2 Nov 2, 2023
0.5.0 Jul 28, 2023
0.4.1 Mar 20, 2023
0.4.0 Jan 26, 2022

#40 in Debugging

Download history 14907/week @ 2023-12-23 20771/week @ 2023-12-30 16932/week @ 2024-01-06 20347/week @ 2024-01-13 15424/week @ 2024-01-20 16340/week @ 2024-01-27 7750/week @ 2024-02-03 1403/week @ 2024-02-10 6726/week @ 2024-02-17 6764/week @ 2024-02-24 5344/week @ 2024-03-02 7338/week @ 2024-03-09 6599/week @ 2024-03-16 7872/week @ 2024-03-23 4014/week @ 2024-03-30 7793/week @ 2024-04-06

27,772 downloads per month
Used in 9 crates (8 directly)

Apache-2.0

175KB
3.5K SLoC

minitrace: Extremely fast tracing library for Rust

Crates.io Documentation CI Status Coverage License


minitrace is a tracing library 10~100x faster than others:

benchmark

Features:

Resources

Getting Started

In Libraries

Libraries should include minitrace as a dependency without enabling any extra features.

[dependencies]
minitrace = "0.6"

Add a trace attribute to the function you want to trace. In this example, a SpanRecord will be collected every time the function is called, if a tracing context is set up by the caller.

#[minitrace::trace]
pub fn send_request(req: HttpRequest) -> Result<(), Error> {
    // ...
}

Libraries are able to set up an individual tracing context, regardless of whether the caller has set up a tracing context or not. This can be achieved by using Span::root() to start a new trace and Span::set_local_parent() to set up a local context for the current thread.

The full_name!() macro can detect the function's full name, which is used as the name of the root span.

use minitrace::prelude::*;

pub fn send_request(req: HttpRequest) -> Result<(), Error> {
    let root = Span::root(full_name!(), SpanContext::random());
    let _guard = root.set_local_parent();

    // ...
}

In Applications

Applications should include minitrace as a dependency with the enable feature set. To disable minitrace statically, simply remove the enable feature.

[dependencies]
minitrace = { version = "0.6", features = ["enable"] }

Applications should initialize a Reporter implementation early in the program's runtime. Span records generated before the reporter is initialized will be ignored. Before terminating, flush() should be called to ensure all collected span records are reported.

When the root span is dropped, all of its children spans and itself will be reported at once. Since that, it's recommended to create root spans for short tasks, such as handling a request, just like the example below. Otherwise, an endingless trace will never be reported.

use minitrace::collector::Config;
use minitrace::collector::ConsoleReporter;
use minitrace::prelude::*;

fn main() {
    minitrace::set_reporter(ConsoleReporter, Config::default());

    loop {
        let root = Span::root("worker-loop", SpanContext::random());
        let _guard = root.set_local_parent();

        handle_request();
    }

    minitrace::flush();
}

Benchmarks

By different architectures:

Benchmark result by architecture

x86-64 (Intel Broadwell) x86-64 (Intel Skylake) x86-64 (AMD Zen) ARM (AWS Graviton2)
tokio-tracing 124x slower 33x slower 36x slower 29x slower
rustracing 45x slower 10x slower 11x slower 9x slower
minitrace (baseline) 1x (3.4us) 1x (3.2us) 1x (3.8us) 1x (4.2us)

By creating different number of spans:

Benchmark result by number of spans

1 span 10 spans 100 spans 1000 spans
tokio-tracing 19x slower 61x slower 124x slower 151x slower
rustracing 13x slower 26x slower 45x slower 55x slower
minitrace (baseline) 1x (0.4us) 1x (0.8us) 1x (3.4us) 1x (27.8us)

Detailed results are available in etc/benchmark-result.

Projects using minitrace

Feel free to open a PR and add your projects here:

  • TiKV: A distributed transactional key-value database
  • Conductor: Open-source GraphQL Gateway
  • Apache OpenDAL: A data access layer for various storage
  • Databend: Cost-Effective alternative to Snowflake

FAQ

Why is minitrace so fast?

There are some articles posted by the maintainer of minitrace:

What is library-level tracing?

Library-level tracing refers to the capability of incorporating tracing capabilities directly within libraries, as opposed to restricting them to application-level or system-level tracing.

Tracing can introduce overhead to a program's execution. While this is generally acceptable at the application level, where the added overhead is often insignificant compared to the overall execution time, it can be more problematic at the library level. Here, functions may be invoked frequently or performance may be critical, and the overhead from tracing can become substantial. As a result, tracing libraries not designed with speed and efficiency in mind may not be suitable for library-level tracing.

In the realm of the minitrace library, library-level tracing is engineered to be fast and lightweight, resulting in zero overhead when it's not activated. This makes minitrace an excellent choice for use in performance-sensitive applications, and it can be seamlessly integrated into libraries in a similar fashion to the log crate, something other tracing libraries may not offer.

How does minitrace differ from other tracing libraries?

While many tracing libraries aim for extensive features, minitrace prioritizes performance and simplicity.

For example, minitrace doesn't introduce new logging macros, e.g. info!() or error!(), but seamlessly integrates with the log crate. This allows you to use existing logging macros and dependencies, with logs automatically attached to the current tracing span.

Will minitrace incorporate 'level' for spans?

The concept of 'level' may not be an optimal feature for tracing systems. While tokio-tracing incorporates this feature, the underlying motivation for having levels in a span primarily revolves around performance. More specifically, it relates to the performance implications of tracing elements that are not of interest. However, tracing differs from logging in two key aspects:

  1. Disregarding a low-level span might inadvertently discard a high-level child span.
  2. The process of filtering, or 'level' as it's often called, in a tracing system should be applied to a trace as a whole rather than individual spans within a trace.

In this context, minitrace offers a more efficient solution by filtering out entire traces that are not of interest through its unique tail-sampling design. Therefore, the concept of 'level', borrowed directly from logging systems, may not be suitable for minitrace.

Will minitrace support OpenTelemetry feature 'X'?

minitrace is focused on high performance tracing only. You can open an issue for the missing tracing features you want to have.

Note that we always prioritize performance over features, so that not all tracing feature requests may be accepted.

What's the status of this library?

API Unstable: The API is not stabilized yet, may be changed in the future.

Code base Tested: minitrace has been tested with high coverage. However, applications utilizing minitrace have not been widely deployed, so that minitrace is currently NOT regarded as battle-tested.

Dependencies

~2.7–9.5MB
~70K SLoC