Crate metrics

Source
Expand description

A lightweight metrics facade.

The metrics crate provides a single metrics API that abstracts over the actual metrics implementation. Libraries can use the metrics API provided by this crate, and the consumer of those libraries can choose the metrics implementation that is most suitable for its use case.

§Overview

metrics exposes two main concepts: emitting a metric, and recording it.

§Metric types, or kinds

This crate supports three fundamental metric types, or kinds: counters, gauges, and histograms.

§Counters

A counter is a cumulative metric that represents a monotonically increasing value which can only be increased or be reset to zero on restart. For example, you might use a counter to represent the number of operations performed, or the number of errors that have occurred.

Counters are unsigned 64-bit integers.

If you have a value that goes up and down over time, consider using a gauge.

§Gauges

A gauge is a metric that can go up and down, arbitrarily, over time.

Gauges are typically used for measured, external values, such as temperature, throughput, or things like current memory usage. Even if the value is monotonically increasing, but there is no way to store the delta in order to properly figure out how much to increment by, then a gauge might be a suitable choice.

Gauges support two modes: incremental updates, or absolute updates. This allows callers to use them for external measurements – where no delta can be computed – as well as internal measurements.

Gauges are floating-point 64-bit numbers.

§Histograms

A histogram stores an arbitrary number of observations of a specific measurement and provides statistical analysis over the observed values. Typically, measurements such as request latency are recorded with histograms: a specific action that is repeated over and over which can have a varying result each time.

Histograms are used to explore the distribution of values, allowing a caller to understand the modalities of the distribution, such as whether or not all values are grouped close together, or spread evenly, or even whether or not there are multiple groupings or clusters.

Colloquially, histograms are usually associated with percentiles, although by definition, they specifically deal with bucketed or binned values: how many values fell within 0-10, how many fell within 11-20, and so on and so forth. Percentiles, commonly associated with “summaries”, deal with understanding how much of a distribution falls below or at a particular percentage of that distribution: 50% of requests are faster than 500ms, 99% of requests are faster than 2450ms, and so on and so forth.

While we use the term “histogram” in metrics, we enforce no particular usage of true histograms or summaries. The choice of output is based entirely on the exporter being used to ship your metric data out of your application. For example, if you’re using metrics-exporter-prometheus, Prometheus supports both histograms and summaries, and the exporter can be configured to output our “histogram” data as either. Other exporters may choose to stick to using summaries, as is traditional, in order to generate percentile data.

Histograms take floating-point 64-bit numbers.

§Emission

Metrics are emitted by utilizing the emission methods. There is a macro for registering and returning a handle for each fundamental metric type:

Additionally, metrics can be described – setting either the unit of measure or long-form description – by using the describe_* macros:

In order to register or emit a metric, you need a way to record these events, which is where Recorder comes into play.

§Recording

The Recorder trait defines the interface between the registration/emission macros, and exporters, which is how we refer to concrete implementations of Recorder. The trait defines what the exporters are doing – recording – but ultimately exporters are sending data from your application to somewhere else: whether it be a third-party service or logging via standard out. It’s “exporting” the metric data out of your application.

Each metric type is usually reserved for a specific type of use case, whether it be tracking a single value or allowing the summation of multiple values, and the respective macros elaborate more on the usage and invariants provided by each.

§Getting Started

§In libraries

Libraries need only include the metrics crate to emit metrics. When an executable installs a recorder, all included crates which emitting metrics will now emit their metrics to that record, which allows library authors to seamless emit their own metrics without knowing or caring which exporter implementation is chosen, or even if one is installed.

In cases where no global recorder is installed, a “noop” recorder lives in its place, which has an incredibly very low overhead: an atomic load and comparison. Libraries can safely instrument their code without fear of ruining baseline performance.

By default, a “noop” recorder is present so that the macros can work even if no exporter has been installed. This recorder has extremely low overhead – a relaxed load and conditional – and so, practically speaking, the overhead when no exporter is installed is extremely low. You can safely instrument applications knowing that you won’t pay a heavy performance cost even if you’re not shipping metrics.

§Examples

use metrics::{counter, histogram};

pub fn process(query: &str) -> u64 {
    let start = Instant::now();
    let row_count = run_query(query);
    let delta = start.elapsed();

    histogram!("process.query_time").record(delta);
    counter!("process.query_row_count").increment(row_count);

    row_count
}

§In executables

Executables, which themselves can emit their own metrics, are intended to install a global recorder so that metrics can actually be recorded and exported somewhere.

Initialization of the global recorder isn’t required for macros to function, but any metrics emitted before a global recorder is installed will not be recorded, so initialization and installation of an exporter should happen as early as possible in the application lifecycle.

§Warning

The metrics system may only be initialized once.

For most use cases, you’ll be using an off-the-shelf exporter implementation that hooks up to an existing metrics collection system, or interacts with the existing systems/processes that you use.

Out of the box, some exporter implementations are available for you to use:

You can also implement your own recorder if a suitable one doesn’t already exist.

§Development

The primary interface with metrics is through the Recorder trait, which is the connection between the user-facing emission macros – counter!, and so on – and the actual logic for handling those metrics and doing something with them, like logging them to the console or sending them to a remote metrics system.

§Keys

All metrics are, in essence, the combination of a metric type and metric identifier, such as a histogram called “response_latency”. You could conceivably have multiple metrics with the same name, so long as they are of different types.

As the types are enforced/limited by the Recorder trait itself, the remaining piece is the identifier, which we handle by using Key. Keys hold both the metric name, and potentially, labels related to the metric. The metric name and labels are always string values.

Internally, metrics uses a clone-on-write “smart pointer” for these values to optimize cases where the values are static strings, which can provide significant performance benefits. These smart pointers can also hold owned String values, though, so users can mix and match static strings and owned strings without issue.

Two Key objects can be checked for equality and considered to point to the same metric if they are equal. Equality checks both the name of the key and the labels of a key. Labels are not sorted prior to checking for equality, but insertion order is maintained, so any Key constructed from the same set of labels in the same order should be equal.

It is an implementation detail if a recorder wishes to do an deeper equality check that ignores the order of labels, but practically speaking, metric emission, and thus labels, should be fixed in ordering in nearly all cases, and so it typically is not a problem.

§Registration

Recorders must handle the “registration” of a metric.

In practice, registration solves two potential problems: providing metadata for a metric, and creating an entry for a metric even though it has not been emitted yet.

Callers may wish to provide a human-readable description of what the metric is, or provide the units the metrics uses. Additionally, users may wish to register their metrics so that they show up in the output of the installed exporter even if the metrics have yet to be emitted. This allows callers to ensure the metrics output is stable, or allows them to expose all of the potential metrics a system has to offer, again, even if they have not all yet been emitted.

As you can see from the trait, the registration methods treats the metadata as optional, and the macros allow users to mix and match whichever fields they want to provide.

When a metric is registered, the expectation is that it will show up in output with a default value, so, for example, a counter should be initialized to zero, a histogram would have no values, and so on.

§Metadata

When registering a metric, metadata can be provided to further describe the metric, in particular about where in the system it originates from and how verbose it is. This metadata emulates much of the same metadata as tracing, as it is intended to be used in a similar way: to provide the ability to filter metrics in a more granular way.

Metadata provides three main pieces of information: the verbosity of the metric (level), the part of the system it originates from (target), and the Rust module it originates from (module path).

For example, an application may wish to collect high-cardinality metrics, such as telemetry about a feature, including the customers using it. Tracking customer usage could mean having a tag with many possible values, and submitting these metrics to the configured downstream system could be costly or computationally expensive.

By setting these metrics to a verbosity level of DEBUG, these metrics could potentially be filtered out at the recorder level, without having to change the application code or manually decide, at the callsite, whether or not to emit the metric.

Metadata is exporter-specific, and may be ignored entirely. See the documentation of the specific exporter being used for more information on how metadata is utilized, if at all.

§Emission

Likewise, recorders must handle the emission of metrics as well.

Comparatively speaking, emission is not too different from registration: you have access to the same Key as well as the value being emitted.

For recorders which temporarily buffer or hold on to values before exporting, a typical approach would be to utilize atomic variables for the storage. For counters and gauges, this can be done simply by using types like AtomicU64. For histograms, this can be slightly tricky as you must hold on to all of the distinct values. In our helper crate, metrics-util, we’ve provided a type called AtomicBucket. For exporters that will want to get all of the current values in a batch, while clearing the bucket so that values aren’t processed again, AtomicBucket provides a simple interface to do so, as well as optimized performance on both the insertion and read side.

Combined together, exporter authors can use Handle, also from the metrics-util crate, which provides a consolidated type for holding metric data. These types, and many more from the metrics-util crate, form the basis of typical exporter behavior and have been exposed to help you quickly build a new exporter.

§Installing recorders

Recorders, also referred to as exporters, must be “installed” such that the emission macros can access them. As users of metrics, you’ll typically see exporters provide methods to install themselves that hide the nitty gritty details. These methods will usually be aptly named, such as install.

However, at a low level, this can happen in one of two ways: installing a recorder globally, or temporarily using it locally.

§Global recorder

The global recorder is the recorder that the macros use by default. It is stored in a static variable accessible by all portions of the compiled application, including dependencies. This is what allows us to provide the same “initialize once, benefit everywhere” behavior that users are familiar with from other telemetry crates like tracing and log.

Only one global recorder can be installed in the lifetime of the process. If a global recorder has already been installed, it cannot be replaced: this is due to the fact that once installed, the recorder is “leaked” so that a static reference can be obtained to it and used by subsequent calls to the emission macros, and any downstream crates.

§Local recorder

In many scenarios, such as in unit tests, you may wish to temporarily set a recorder to influence all calls to the emission macros within a specific section of code, without influencing other areas of the code, or being limited by the constraints of only one global recorder being allowed.

with_local_recorder allows you to do this by changing the recorder used by the emission macros for the duration of a given closure. While in that closure, the given recorder will act as if it was the global recorder for the current thread. Once the closure returns, the true global recorder takes priority again for the current thread.

Modules§

  • Atomic types used for metrics.

Macros§

Structs§

Enums§

Traits§

Functions§

Type Aliases§