Expand description

The Layer trait, a composable abstraction for building Subscribers.

The Subscriber trait in tracing-core represents the complete set of functionality required to consume tracing instrumentation. This means that a single Subscriber instance is a self-contained implementation of a complete strategy for collecting traces; but it also means that the Subscriber trait cannot easily be composed with other Subscribers.

In particular, Subscribers are responsible for generating span IDs and assigning them to spans. Since these IDs must uniquely identify a span within the context of the current trace, this means that there may only be a single Subscriber for a given thread at any point in time — otherwise, there would be no authoritative source of span IDs.

On the other hand, the majority of the Subscriber trait’s functionality is composable: any number of subscribers may observe events, span entry and exit, and so on, provided that there is a single authoritative source of span IDs. The Layer trait represents this composable subset of the Subscriber behavior; it can observe events and spans, but does not assign IDs.

Composing Layers

Since a Layer does not implement a complete strategy for collecting traces, it must be composed with a Subscriber in order to be used. The Layer trait is generic over a type parameter (called S in the trait definition), representing the types of Subscriber they can be composed with. Thus, a Layer may be implemented that will only compose with a particular Subscriber implementation, or additional trait bounds may be added to constrain what types implementing Subscriber a Layer can wrap.

Layers may be added to a Subscriber by using the SubscriberExt::with method, which is provided by tracing-subscriber’s prelude. This method returns a Layered struct that implements Subscriber by composing the Layer with the Subscriber.

For example:

use tracing_subscriber::Layer;
use tracing_subscriber::prelude::*;
use tracing::Subscriber;

pub struct MyLayer {
    // ...
}

impl<S: Subscriber> Layer<S> for MyLayer {
    // ...
}

pub struct MySubscriber {
    // ...
}

impl Subscriber for MySubscriber {
    // ...
}

let subscriber = MySubscriber::new()
    .with(MyLayer::new());

tracing::subscriber::set_global_default(subscriber);

Multiple Layers may be composed in the same manner:

pub struct MyOtherLayer {
    // ...
}

impl<S: Subscriber> Layer<S> for MyOtherLayer {
    // ...
}

pub struct MyThirdLayer {
    // ...
}

impl<S: Subscriber> Layer<S> for MyThirdLayer {
    // ...
}
}

let subscriber = MySubscriber::new()
    .with(MyLayer::new())
    .with(MyOtherLayer::new())
    .with(MyThirdLayer::new());

tracing::subscriber::set_global_default(subscriber);

The Layer::with_subscriber constructs the Layered type from a Layer and Subscriber, and is called by SubscriberExt::with. In general, it is more idiomatic to use SubscriberExt::with, and treat Layer::with_subscriber as an implementation detail, as with_subscriber calls must be nested, leading to less clear code for the reader.

Runtime Configuration With Layers

In some cases, a particular Layer may be enabled or disabled based on runtime configuration. This can introduce challenges, because the type of a layered Subscriber depends on which layers are added to it: if an if or match expression adds some Layer implementation in one branch, and other layers in another, the Subscriber values returned by those branches will have different types. For example, the following will not work:

use std::fs::File;
use tracing_subscriber::{Registry, prelude::*};

let stdout_log = tracing_subscriber::fmt::layer().pretty();
let subscriber = Registry::default().with(stdout_log);

// The compile error will occur here because the if and else
// branches have different (and therefore incompatible) types.
let subscriber = if cfg.is_prod {
    let file = File::create(cfg.path)?;
    let layer = tracing_subscriber::fmt::layer()
        .json()
        .with_writer(Arc::new(file));
    layer.with(subscriber)
} else {
    layer
};

tracing::subscriber::set_global_default(subscriber)
    .expect("Unable to set global subscriber");

However, a Layer wrapped in an Option also implements the Layer trait. This allows individual layers to be enabled or disabled at runtime while always producing a Subscriber of the same type. For example:

use std::fs::File;
use tracing_subscriber::{Registry, prelude::*};

let stdout_log = tracing_subscriber::fmt::layer().pretty();
let subscriber = Registry::default().with(stdout_log);

// if `cfg.is_prod` is true, also log JSON-formatted logs to a file.
let json_log = if cfg.is_prod {
    let file = File::create(cfg.path)?;
    let json_log = tracing_subscriber::fmt::layer()
        .json()
        .with_writer(file);
    Some(json_log)
} else {
    None
};

// If `cfg.is_prod` is false, then `json` will be `None`, and this layer
// will do nothing. However, the subscriber will still have the same type
// regardless of whether the `Option`'s value is `None` or `Some`.
let subscriber = subscriber.with(json_log);

tracing::subscriber::set_global_default(subscriber)
   .expect("Unable to set global subscriber");

If a Layer may be one of several different types, note that Box<dyn Layer<S> + Send + Sync> implements Layer. This may be used to erase the type of a Layer.

For example, a function that configures a Layer to log to one of several outputs might return a Box<dyn Layer<S> + Send + Sync + 'static>:

use tracing_subscriber::{
    Layer,
    registry::LookupSpan,
    prelude::*,
};
use std::{path::PathBuf, fs::File, io};

/// Configures whether logs are emitted to a file, to stdout, or to stderr.
pub enum LogConfig {
    File(PathBuf),
    Stdout,
    Stderr,
}

impl LogConfig {
    pub fn layer<S>(self) -> Box<dyn Layer<S> + Send + Sync + 'static>
    where
        S: tracing_core::Subscriber,
        for<'a> S: LookupSpan<'a>,
    {
        // Shared configuration regardless of where logs are output to.
        let fmt = tracing_subscriber::fmt::layer()
            .with_target(true)
            .with_thread_names(true);

        // Configure the writer based on the desired log target:
        match self {
            LogConfig::File(path) => {
                let file = File::create(path).expect("failed to create log file");
                Box::new(fmt.with_writer(file))
            },
            LogConfig::Stdout => Box::new(fmt.with_writer(io::stdout)),
            LogConfig::Stderr => Box::new(fmt.with_writer(io::stderr)),
        }
    }
}

let config = LogConfig::Stdout;
tracing_subscriber::registry()
    .with(config.layer())
    .init();

The Layer::boxed method is provided to make boxing a Layer more convenient, but Box::new may be used as well.

When the number of Layers varies at runtime, note that a Vec<L> where L: Layer also implements Layer. This can be used to add a variable number of Layers to a Subscriber:

use tracing_subscriber::{Layer, prelude::*};
struct MyLayer {
    // ...
}

impl<S: tracing_core::Subscriber> Layer<S> for MyLayer {
    // ...
}

/// Returns how many layers we need
fn how_many_layers() -> usize {
    // ...
}

// Create a variable-length `Vec` of layers
let mut layers = Vec::new();
for _ in 0..how_many_layers() {
    layers.push(MyLayer::new());
}

tracing_subscriber::registry()
    .with(layers)
    .init();

If a variable number of Layer is needed and those Layers have different types, a Vec of boxed Layer trait objects may be used. For example:

use tracing_subscriber::{filter::LevelFilter, Layer, prelude::*};
use std::fs::File;
struct Config {
    enable_log_file: bool,
    enable_stdout: bool,
    enable_stderr: bool,
    // ...
}

let cfg = Config::from_config_file()?;

// Based on our dynamically loaded config file, create any number of layers:
let mut layers = Vec::new();

if cfg.enable_log_file {
    let file = File::create("myapp.log")?;
    let layer = tracing_subscriber::fmt::layer()
        .with_thread_names(true)
        .with_target(true)
        .json()
        .with_writer(file)
        // Box the layer as a type-erased trait object, so that it can
        // be pushed to the `Vec`.
        .boxed();
    layers.push(layer);
}

if cfg.enable_stdout {
    let layer = tracing_subscriber::fmt::layer()
        .pretty()
        .with_filter(LevelFilter::INFO)
        // Box the layer as a type-erased trait object, so that it can
        // be pushed to the `Vec`.
        .boxed();
    layers.push(layer);
}

if cfg.enable_stdout {
    let layer = tracing_subscriber::fmt::layer()
        .with_target(false)
        .with_filter(LevelFilter::WARN)
        // Box the layer as a type-erased trait object, so that it can
        // be pushed to the `Vec`.
        .boxed();
    layers.push(layer);
}

tracing_subscriber::registry()
    .with(layers)
    .init();

Finally, if the number of layers changes at runtime, a Vec of subscribers can be used alongside the reload module to add or remove subscribers dynamically at runtime.

Recording Traces

The Layer trait defines a set of methods for consuming notifications from tracing instrumentation, which are generally equivalent to the similarly named methods on Subscriber. Unlike Subscriber, the methods on Layer are additionally passed a Context type, which exposes additional information provided by the wrapped subscriber (such as the current span) to the layer.

Filtering with Layers

As well as strategies for handling trace events, the Layer trait may also be used to represent composable filters. This allows the determination of what spans and events should be recorded to be decoupled from how they are recorded: a filtering layer can be applied to other layers or subscribers. Layers can be used to implement global filtering, where a Layer provides a filtering strategy for the entire subscriber. Additionally, individual recording Layers or sets of Layers may be combined with per-layer filters that control what spans and events are recorded by those layers.

Global Filtering

A Layer that implements a filtering strategy should override the register_callsite and/or enabled methods. It may also choose to implement methods such as on_enter, if it wishes to filter trace events based on the current span context.

Note that the Layer::register_callsite and Layer::enabled methods determine whether a span or event is enabled globally. Thus, they should not be used to indicate whether an individual layer wishes to record a particular span or event. Instead, if a layer is only interested in a subset of trace data, but does not wish to disable other spans and events for the rest of the layer stack should ignore those spans and events in its notification methods.

The filtering methods on a stack of Layers are evaluated in a top-down order, starting with the outermost Layer and ending with the wrapped Subscriber. If any layer returns false from its enabled method, or Interest::never() from its register_callsite method, filter evaluation will short-circuit and the span or event will be disabled.

Enabling Interest

Whenever an tracing event (or span) is emitted, it goes through a number of steps to determine how and how much it should be processed. The earlier an event is disabled, the less work has to be done to process the event, so Layers that implement filtering should attempt to disable unwanted events as early as possible. In order, each event checks:

  • register_callsite, once per callsite (roughly: once per time that event! or span! is written in the source code; this is cached at the callsite). See Subscriber::register_callsite and tracing_core::callsite for a summary of how this behaves.
  • enabled, once per emitted event (roughly: once per time that event! or span! is executed), and only if register_callsite regesters an Interest::sometimes. This is the main customization point to globally filter events based on their Metadata. If an event can be disabled based only on Metadata, it should be, as this allows the construction of the actual Event/Span to be skipped.
  • For events only (and not spans), event_enabled is called just before processing the event. This gives layers one last chance to say that an event should be filtered out, now that the event’s fields are known.

Per-Layer Filtering

Note: per-layer filtering APIs currently require the "registry" crate feature flag to be enabled.

Sometimes, it may be desirable for one Layer to record a particular subset of spans and events, while a different subset of spans and events are recorded by other Layers. For example:

  • A layer that records metrics may wish to observe only events including particular tracked values, while a logging layer ignores those events.
  • If recording a distributed trace is expensive, it might be desirable to only send spans with INFO and lower verbosity to the distributed tracing system, while logging more verbose spans to a file.
  • Spans and events with a particular target might be recorded differently from others, such as by generating an HTTP access log from a span that tracks the lifetime of an HTTP request.

The Filter trait is used to control what spans and events are observed by an individual Layer, while still allowing other Layers to potentially record them. The Layer::with_filter method combines a Layer with a Filter, returning a Filtered layer.

This crate’s filter module provides a number of types which implement the Filter trait, such as LevelFilter, Targets, and FilterFn. These Filters provide ready-made implementations of common forms of filtering. For custom filtering policies, the FilterFn and DynFilterFn types allow implementing a Filter with a closure or function pointer. In addition, when more control is required, the Filter trait may also be implemented for user-defined types.

     Warning: Currently, the 
     Registry type defined in this crate is the only root
     Subscriber capable of supporting Layers with
     per-layer filters. In the future, new APIs will be added to allow other
     root Subscribers to support per-layer filters.
 

For example, to generate an HTTP access log based on spans with the http_access target, while logging other spans and events to standard out, a Filter can be added to the access log layer:

use tracing_subscriber::{filter, prelude::*};

// Generates an HTTP access log.
let access_log = // ...

// Add a filter to the access log layer so that it only observes
// spans and events with the `http_access` target.
let access_log = access_log.with_filter(filter::filter_fn(|metadata| {
    // Returns `true` if and only if the span or event's target is
    // "http_access".
    metadata.target() == "http_access"
}));

// A general-purpose logging layer.
let fmt_layer = tracing_subscriber::fmt::layer();

// Build a subscriber that combines the access log and stdout log
// layers.
tracing_subscriber::registry()
    .with(fmt_layer)
    .with(access_log)
    .init();

Multiple layers can have their own, separate per-layer filters. A span or event will be recorded if it is enabled by any per-layer filter, but it will be skipped by the layers whose filters did not enable it. Building on the previous example:

use tracing_subscriber::{filter::{filter_fn, LevelFilter}, prelude::*};

let access_log = // ...
let fmt_layer = tracing_subscriber::fmt::layer();

tracing_subscriber::registry()
    // Add the filter for the "http_access" target to the access
    // log layer, like before.
    .with(access_log.with_filter(filter_fn(|metadata| {
        metadata.target() == "http_access"
    })))
    // Add a filter for spans and events with the INFO level
    // and below to the logging layer.
    .with(fmt_layer.with_filter(LevelFilter::INFO))
    .init();

// Neither layer will observe this event
tracing::debug!(does_anyone_care = false, "a tree fell in the forest");

// This event will be observed by the logging layer, but not
// by the access log layer.
tracing::warn!(dose_roentgen = %3.8, "not great, but not terrible");

// This event will be observed only by the access log layer.
tracing::trace!(target: "http_access", "HTTP request started");

// Both layers will observe this event.
tracing::error!(target: "http_access", "HTTP request failed with a very bad error!");

A per-layer filter can be applied to multiple Layers at a time, by combining them into a Layered layer using Layer::and_then, and then calling Layer::with_filter on the resulting Layered layer.

Consider the following:

  • layer_a and layer_b, which should only receive spans and events at the INFO level and above.
  • A third layer, layer_c, which should receive spans and events at the DEBUG level as well. The layers and filters would be composed thusly:
use tracing_subscriber::{filter::LevelFilter, prelude::*};

let layer_a = // ...
let layer_b =  // ...
let layer_c =  // ...

let info_layers = layer_a
    // Combine `layer_a` and `layer_b` into a `Layered` layer:
    .and_then(layer_b)
    // ...and then add an `INFO` `LevelFilter` to that layer:
    .with_filter(LevelFilter::INFO);

tracing_subscriber::registry()
    // Add `layer_c` with a `DEBUG` filter.
    .with(layer_c.with_filter(LevelFilter::DEBUG))
    .with(info_layers)
    .init();

If a Filtered Layer is combined with another Layer Layer::and_then, and a filter is added to the Layered layer, that layer will be filtered by both the inner filter and the outer filter. Only spans and events that are enabled by both filters will be observed by that layer. This can be used to implement complex filtering trees.

As an example, consider the following constraints:

  • Suppose that a particular target is used to indicate events that should be counted as part of a metrics system, which should be only observed by a layer that collects metrics.
  • A log of high-priority events (INFO and above) should be logged to stdout, while more verbose events should be logged to a debugging log file.
  • Metrics-focused events should not be included in either log output.

In that case, it is possible to apply a filter to both logging layers to exclude the metrics events, while additionally adding a LevelFilter to the stdout log:

use tracing_subscriber::{filter, prelude::*};
use std::{fs::File, sync::Arc};

// A layer that logs events to stdout using the human-readable "pretty"
// format.
let stdout_log = tracing_subscriber::fmt::layer()
    .pretty();

// A layer that logs events to a file.
let file = File::create("debug.log")?;
let debug_log = tracing_subscriber::fmt::layer()
    .with_writer(Arc::new(file));

// A layer that collects metrics using specific events.
let metrics_layer = /* ... */ filter::LevelFilter::INFO;

tracing_subscriber::registry()
    .with(
        stdout_log
            // Add an `INFO` filter to the stdout logging layer
            .with_filter(filter::LevelFilter::INFO)
            // Combine the filtered `stdout_log` layer with the
            // `debug_log` layer, producing a new `Layered` layer.
            .and_then(debug_log)
            // Add a filter to *both* layers that rejects spans and
            // events whose targets start with `metrics`.
            .with_filter(filter::filter_fn(|metadata| {
                !metadata.target().starts_with("metrics")
            }))
    )
    .with(
        // Add a filter to the metrics label that *only* enables
        // events whose targets start with `metrics`.
        metrics_layer.with_filter(filter::filter_fn(|metadata| {
            metadata.target().starts_with("metrics")
        }))
    )
    .init();

// This event will *only* be recorded by the metrics layer.
tracing::info!(target: "metrics::cool_stuff_count", value = 42);

// This event will only be seen by the debug log file layer:
tracing::debug!("this is a message, and part of a system of messages");

// This event will be seen by both the stdout log layer *and*
// the debug log file layer, but not by the metrics layer.
tracing::warn!("the message is a warning about danger!");

Structs

Traits

  • A per-Layer filter that determines whether a span or event is enabled for an individual layer.
  • A composable handler for tracing events.
  • Extension trait adding a with(Layer) combinator to Subscribers.