Attribute Macro tracing::instrument

source ·
#[instrument]
Expand description

Instruments a function to create and enter a tracing span every time the function is called.

Unless overriden, a span with the INFO level will be generated. The generated span’s name will be the name of the function. By default, all arguments to the function are included as fields on the span. Arguments that are tracing primitive types implementing the Value trait will be recorded as fields of that type. Types which do not implement Value will be recorded using [std::fmt::Debug].

Overriding Span Attributes

To change the name of the generated span, add a name argument to the #[instrument] macro, followed by an equals sign and a string literal. For example:


// The generated span's name will be "my_span" rather than "my_function".
#[instrument(name = "my_span")]
pub fn my_function() {
    // ... do something incredibly interesting and important ...
}

To override the target of the generated span, add a target argument to the #[instrument] macro, followed by an equals sign and a string literal for the new target. The module path is still recorded separately. For example:

pub mod my_module {
    // The generated span's target will be "my_crate::some_special_target",
    // rather than "my_crate::my_module".
    #[instrument(target = "my_crate::some_special_target")]
    pub fn my_function() {
        // ... all kinds of neat code in here ...
    }
}

Finally, to override the level of the generated span, add a level argument, followed by an equals sign and a string literal with the name of the desired level. Level names are not case sensitive. For example:

// The span's level will be TRACE rather than INFO.
#[instrument(level = "trace")]
pub fn my_function() {
    // ... I have written a truly marvelous implementation of this function,
    // which this example is too narrow to contain ...
}

Skipping Fields

To skip recording one or more arguments to a function or method, pass the argument’s name inside the skip() argument on the #[instrument] macro. This can be used when an argument to an instrumented function does not implement fmt::Debug, or to exclude an argument with a verbose or costly Debug implementation. Note that:

  • multiple argument names can be passed to skip.
  • arguments passed to skip do not need to implement fmt::Debug.

You can also use skip_all to skip all arguments.

Examples

// This type doesn't implement `fmt::Debug`!
struct NonDebug;

// `arg` will be recorded, while `non_debug` will not.
#[instrument(skip(non_debug))]
fn my_function(arg: usize, non_debug: NonDebug) {
    // ...
}

// These arguments are huge
#[instrument(skip_all)]
fn my_big_data_function(large: Vec<u8>, also_large: HashMap<String, String>) {
    // ...
}

Skipping the self parameter:

#[derive(Debug)]
struct MyType {
   data: Vec<u8>, // Suppose this buffer is often quite long...
}

impl MyType {
    // Suppose we don't want to print an entire kilobyte of `data`
    // every time this is called...
    #[instrument(skip(self))]
    pub fn my_method(&mut self, an_interesting_argument: usize) {
         // ... do something (hopefully, using all that `data`!)
    }
}

Adding Fields

Additional fields (key-value pairs with arbitrary data) may be added to the generated span using the fields argument on the #[instrument] macro. Any Rust expression can be used as a field value in this manner. These expressions will be evaluated at the beginning of the function’s body, so arguments to the function may be used in these expressions. Field names may also be specified without values. Doing so will result in an empty field whose value may be recorded later within the function body.

This supports the same field syntax as the span! and event! macros.

Note that overlap between the names of fields and (non-skipped) arguments will result in a compile error.

Examples

Adding a new field based on the value of an argument:


// This will record a field named "i" with the value of `i` *and* a field
// named "next" with the value of `i` + 1.
#[instrument(fields(next = i + 1))]
pub fn my_function(i: usize) {
    // ...
}

Recording specific properties of a struct as their own fields:


// This will record the request's URI and HTTP method as their own separate
// fields.
#[instrument(fields(http.uri = req.uri(), http.method = req.method()))]
pub fn handle_request<B>(req: http::Request<B>) -> http::Response<B> {
    // ... handle the request ...
}

This can be used in conjunction with skip or skip_all to record only some fields of a struct:

// Remember the struct with the very large `data` field from the earlier
// example? Now it also has a `name`, which we might want to include in
// our span.
#[derive(Debug)]
struct MyType {
   name: &'static str,
   data: Vec<u8>,
}

impl MyType {
    // This will skip the `data` field, but will include `self.name`,
    // formatted using `fmt::Display`.
    #[instrument(skip(self), fields(self.name = %self.name))]
    pub fn my_method(&mut self, an_interesting_argument: usize) {
         // ... do something (hopefully, using all that `data`!)
    }
}

Adding an empty field to be recorded later:


// This function does a very interesting and important mathematical calculation.
// Suppose we want to record both the inputs to the calculation *and* its result...
#[instrument(fields(result))]
pub fn do_calculation(input_1: usize, input_2: usize) -> usize {
    // Rerform the calculation.
    let result = input_1 + input_2;

    // Record the result as part of the current span.
    tracing::Span::current().record("result", &result);

    // Now, the result will also be included on this event!
    tracing::info!("calculation complete!");

    // ... etc ...
}

Examples

Instrumenting a function:

#[instrument]
pub fn my_function(my_arg: usize) {
    // This event will be recorded inside a span named `my_function` with the
    // field `my_arg`.
    tracing::info!("inside my_function!");
    // ...
}

Setting the level for the generated span:

#[instrument(level = "debug")]
pub fn my_function() {
    // ...
}

Overriding the generated span’s name:

#[instrument(name = "my_name")]
pub fn my_function() {
    // ...
}

Overriding the generated span’s target:

#[instrument(target = "my_target")]
pub fn my_function() {
    // ...
}

Overriding the generated span’s parent:

#[instrument(parent = None)]
pub fn my_function() {
    // ...
}
// A struct which owns a span handle.
struct MyStruct
{
    span: tracing::Span
}

impl MyStruct
{
    // Use the struct's `span` field as the parent span
    #[instrument(parent = &self.span, skip(self))]
    fn my_method(&self) {}
}

Specifying follows_from relationships:

#[instrument(follows_from = causes)]
pub fn my_function(causes: &[tracing::Id]) {
    // ...
}

Any expression of type impl IntoIterator<Item = impl Into<Option<Id>>> may be provided to follows_from; e.g.:

#[instrument(follows_from = [cause])]
pub fn my_function(cause: &tracing::span::EnteredSpan) {
    // ...
}

To skip recording an argument, pass the argument’s name to the skip:

struct NonDebug;

#[instrument(skip(non_debug))]
fn my_function(arg: usize, non_debug: NonDebug) {
    // ...
}

To add an additional context to the span, pass key-value pairs to fields:

#[instrument(fields(foo="bar", id=1, show=true))]
fn my_function(arg: usize) {
    // ...
}

Adding the ret argument to #[instrument] will emit an event with the function’s return value when the function returns:

#[instrument(ret)]
fn my_function() -> i32 {
    42
}

The return value event will have the same level as the span generated by #[instrument]. By default, this will be INFO, but if the level is overridden, the event will be at the same level.

Note: if the function returns a Result<T, E>, ret will record returned values if and only if the function returns [Result::Ok].

By default, returned values will be recorded using their [std::fmt::Debug] implementations. If a returned value implements [std::fmt::Display], it can be recorded using its Display implementation instead, by writing ret(Display):

#[instrument(ret(Display))]
fn my_function() -> i32 {
    42
}

If the function returns a Result<T, E> and E implements std::fmt::Display, you can add err or err(Display) to emit error events when the function returns Err:

#[instrument(err)]
fn my_function(arg: usize) -> Result<(), std::io::Error> {
    Ok(())
}

By default, error values will be recorded using their std::fmt::Display implementations. If an error implements std::fmt::Debug, it can be recorded using its Debug implementation instead, by writing err(Debug):

#[instrument(err(Debug))]
fn my_function(arg: usize) -> Result<(), std::io::Error> {
    Ok(())
}

If a target is specified, both the ret and err arguments will emit outputs to the declared target (or the default channel if target is not specified).

The ret and err arguments can be combined in order to record an event if a function returns [Result::Ok] or [Result::Err]:

#[instrument(err, ret)]
fn my_function(arg: usize) -> Result<(), std::io::Error> {
    Ok(())
}

async fns may also be instrumented:

#[instrument]
pub async fn my_function() -> Result<(), ()> {
    // ...
}

It also works with async-trait (a crate that allows defining async functions in traits, something not currently possible in Rust), and hopefully most libraries that exhibit similar behaviors:

use async_trait::async_trait;

#[async_trait]
pub trait Foo {
    async fn foo(&self, arg: usize);
}

#[derive(Debug)]
struct FooImpl(usize);

#[async_trait]
impl Foo for FooImpl {
    #[instrument(fields(value = self.0, tmp = std::any::type_name::<Self>()))]
    async fn foo(&self, arg: usize) {}
}

Note than on async-trait <= 0.1.43, references to the Self type inside the fields argument were only allowed when the instrumented function is a method (i.e., the function receives self as an argument). For example, this used to not work because the instrument function didn’t receive self:

use async_trait::async_trait;

#[async_trait]
pub trait Bar {
    async fn bar();
}

#[derive(Debug)]
struct BarImpl(usize);

#[async_trait]
impl Bar for BarImpl {
    #[instrument(fields(tmp = std::any::type_name::<Self>()))]
    async fn bar() {}
}

Instead, you should manually rewrite any Self types as the type for which you implement the trait: #[instrument(fields(tmp = std::any::type_name::<Bar>()))] (or maybe you can just bump async-trait).