Async tracing capabilities for the standard log crate.
This crate provides extension types and hooks to log to enable asynchronous logging.
When building a synchronous application, log messages can be relied on to always happen in sequence. But unfortunately synchronous applications are rarely capable of utilizating system resources to their full potential.
In contrast, concurrent applications make a lot better use of system resources. But it also means we can no longer rely on log messages to strictly happen in sequence. In order to make sense of logs in asynchronous applications, we need to be able to correlate sequences of events with each other:
a1 -> b1 -> b2 -> a2 -> b3 # raw log stream
a1 -------------> a2 # parsed log stream a
b1 -> b2 -------> b3 # parsed log stream bThe raw log stream contains items for both "a" and "b". With async logging you want to be able to distinguish between the items for "a", and the items from "b".
The goal of async logging is to determine which events happened in sequence inside your code. In
practice this means being able to correlate events with each other past yield points (e.g.
.await), and thread bounds.
The way we do this is by adding the current task ID, and thread ID from where the log is occurring. An whenever a new task is spawned we log the following values:
- The ID of the task from which the new task is spawned (
task_parent_id) - The ID of the new task that's spawned (
task_id) - The current thread ID (
thread_id) - The line from where the task was spawned (
spawn_line, whenRUST_BACKTRACE=1enabled)
With all this information we have all the information to correlate tasks with each other. We know what the parent task was, what the new task is, and log that information together. On the receiving side we can then reconstruct that to create correlations.
A span is a pair of messages. One is emitted at the start of an operation, and the other is emitted at the end of the operation. If we add timestamps to when each message was sent, we're able to determine how long operations take. Or determine which operations never finished.
In async-log each span is annotated with a span_mark message:
span_mark=startmarks the start of a spanspan_mark=endmarks the end of a span
example
runtime::fs::read_to_string, span_mark=start, path=/tmp/foob, task_id=7, thread_id=8
runtime::fs::read_to_string, span_mark=end, path=/tmp/foob, task_id=7, thread_id=8log is Rust's standard log crate. It's incredibly flexible, and was
built with extensibility in mind. Because it's so widely used, being able to extend it allows
us to add tracing data to crates without needing to make any changes to their log! calls.
Structured logging (key-value logging) is currently in the
process of being added to log.
At the time of writing there are no published versions available with even the experimental
features available. So until then we have to add key-value pairs using strings. Once key-value
logging is added to log we'll publish a breaking change, and move over.
The syntax we've chosen to use is foo=bar pairs. Multiple pairs should be delimited using
commas (,). Every pair should come after the first message. An example log looks like this:
a new cat has logged on, name=nori, snacks=alwaysuse async_log::span;
use log::info;
fn setup_logger() {
let logger = femme::pretty::Logger::new();
async_log::Logger::wrap(logger, || 12)
.start(log::LevelFilter::Trace)
.unwrap();
}
fn main() {
setup_logger();
span!("level I", {
let x = "beep";
info!("look at this value, x={}", x);
span!("level II", {
let y = "boop";
info!("another nice value, y={}", y);
})
})
}$ cargo add async-logThis crate uses #![deny(unsafe_code)] to ensure everything is implemented in
100% Safe Rust.
Want to join us? Check out our "Contributing" guide and take a look at some of these issues:
MIT OR Apache-2.0
