Cadence

Cadence Documentation

Macros Documentation

An extensible Statsd client for Rust!

Cadence is a fast and flexible way to emit Statsd metrics from your application.

Features

• Support for emitting counters, timers, histograms, distributions, gauges, meters, and sets to Statsd over UDP (or optionally Unix sockets).
• Support for alternate backends via the MetricSink trait.
• Support for Datadog style metrics tags.
• Macros to simplify common calls to emit metrics
• A simple yet flexible API for sending metrics.

Install

To make use of cadence in your project, add it as a dependency in your Cargo.toml file.

[dependencies]
cadence = "x.y.z"

That's all you need!

Usage

Some examples of how to use Cadence are shown below. The examples start simple and work up to how you should be using Cadence in a production application.

Simple Use

Simple usage of Cadence is shown below. In this example, we just import the client, create an instance that will write to some imaginary metrics server, and send a few metrics.

use std::net::UdpSocket;
use cadence::prelude::*;
use cadence::{StatsdClient, UdpMetricSink, DEFAULT_PORT};
// Create client that will write to the given host over UDP.
//
// Note that you'll probably want to actually handle any errors creating
// the client when you use it for real in your application. We're just
// using .unwrap() here since this is an example!
let host = ("metrics.example.com", DEFAULT_PORT);
let socket = UdpSocket::bind("0.0.0.0:0").unwrap();
let sink = UdpMetricSink::from(host, socket).unwrap();
let client = StatsdClient::from_sink("my.metrics", sink);

// Emit metrics!
client.incr("some.counter");
client.time("some.methodCall", 42);
client.gauge("some.thing", 7);
client.meter("some.value", 5);

Buffered UDP Sink

While sending a metric over UDP is very fast, the overhead of frequent network calls can start to add up. This is especially true if you are writing a high performance application that emits a lot of metrics.

To make sure that metrics aren't interfering with the performance of your application, you may want to use a MetricSink implementation that buffers multiple metrics before sending them in a single network operation. For this, there's BufferedUdpMetricSink. An example of using this sink is given below.

use std::net::UdpSocket;
use cadence::prelude::*;
use cadence::{StatsdClient, BufferedUdpMetricSink, DEFAULT_PORT};

let socket = UdpSocket::bind("0.0.0.0:0").unwrap();
socket.set_nonblocking(true).unwrap();

let host = ("metrics.example.com", DEFAULT_PORT);
let sink = BufferedUdpMetricSink::from(host, socket).unwrap();
let client = StatsdClient::from_sink("my.prefix", sink);

client.count("my.counter.thing", 29);
client.time("my.service.call", 214);
client.incr("some.event");

As you can see, using this buffered UDP sink is no more complicated than using the regular, non-buffered, UDP sink.

The only downside to this sink is that metrics aren't written to the Statsd server until the buffer is full. If you have a busy application that is constantly emitting metrics, this shouldn't be a problem. However, if your application only occasionally emits metrics, this sink might result in the metrics being delayed for a little while until the buffer fills. In this case, it may make sense to use the UdpMetricSink since it does not do any buffering.

Queuing Asynchronous Metric Sink

To make sure emitting metrics doesn't interfere with the performance of your application (even though emitting metrics is generally quite fast), it's probably a good idea to make sure metrics are emitted in in a different thread than your application thread.

To allow you to do this, there is QueuingMetricSink. This sink allows you to wrap any other metric sink and send metrics to it via a queue, as it emits metrics in another thread, asynchronously from the flow of your application.

The requirements for the wrapped metric sink are that it is thread safe, meaning that it implements the Send and Sync traits. If you're using the QueuingMetricSink with another sink from Cadence, you don't need to worry: they are all thread safe.

An example of using the QueuingMetricSink to wrap a buffered UDP metric sink is given below. This is the preferred way to use Cadence in production.

use std::net::UdpSocket;
use cadence::prelude::*;
use cadence::{StatsdClient, QueuingMetricSink, BufferedUdpMetricSink, DEFAULT_PORT};

let socket = UdpSocket::bind("0.0.0.0:0").unwrap();
socket.set_nonblocking(true).unwrap();

let host = ("metrics.example.com", DEFAULT_PORT);
let udp_sink = BufferedUdpMetricSink::from(host, socket).unwrap();
let queuing_sink = QueuingMetricSink::from(udp_sink);
let client = StatsdClient::from_sink("my.prefix", queuing_sink);

client.count("my.counter.thing", 29);
client.time("my.service.call", 214);

In the example above, we use the default constructor for the queuing sink which creates an unbounded queue, with no maximum size, to connect the main thread where the client sends metrics to the background thread in which the wrapped sink is running. If instead, you want to create a bounded queue with a maximum size, you can use the with_capacity constructor. An example of this is given below.

use std::net::UdpSocket;
use cadence::prelude::*;
use cadence::{StatsdClient, QueuingMetricSink, BufferedUdpMetricSink,
              DEFAULT_PORT};

// Queue with a maximum capacity of 128K elements
const QUEUE_SIZE: usize = 128 * 1024;

let socket = UdpSocket::bind("0.0.0.0:0").unwrap();
socket.set_nonblocking(true).unwrap();

let host = ("metrics.example.com", DEFAULT_PORT);
let udp_sink = BufferedUdpMetricSink::from(host, socket).unwrap();
let queuing_sink = QueuingMetricSink::with_capacity(udp_sink, QUEUE_SIZE);
let client = StatsdClient::from_sink("my.prefix", queuing_sink);

client.count("my.counter.thing", 29);
client.time("my.service.call", 214);
client.incr("some.event");

Using a QueuingMetricSink with a capacity set means that when the queue is full, attempts to emit metrics via the StatsdClient will fail. While this is bad, the alternative (if you instead used an unbounded queue) is for unsent metrics to slowly use up more and more memory until your application exhausts all memory.

Using an unbounded queue means that the sending of metrics can absorb slowdowns of sending metrics until your application runs out of memory. Using a bounded queue puts a cap on the amount of memory that sending metrics will use in your application. This is a tradeoff that users of Cadence must decide for themselves.

Use With Tags

Adding tags to metrics is accomplished via the use of each of the _with_tags methods that are part of the Cadence StatsdClient struct. An example of using these methods is given below. Note that tags are an extension to the Statsd protocol and so may not be supported by all servers.

See the Datadog docs for more information.

use cadence::prelude::*;
use cadence::{Metric, StatsdClient, NopMetricSink};

let client = StatsdClient::from_sink("my.prefix", NopMetricSink);

let res = client.count_with_tags("my.counter", 29)
    .with_tag("host", "web03.example.com")
    .with_tag_value("beta-test")
    .try_send();

assert_eq!(
    concat!(
        "my.prefix.my.counter:29|c|#",
        "host:web03.example.com,",
        "beta-test"
    ),
    res.unwrap().as_metric_str()
);

Default Tags

Default tags can be added to a StatsdClient when constructed using the builder. Default tags are added to every metric emitted by the StatsdClient without any extra work after building the client. Note that tags are an extension to the Statsd protocol and so may not be supported by all servers.

See the Datadog docs for more information.

use cadence::prelude::*;
use cadence::{Metric, StatsdClient, NopMetricSink};

let client = StatsdClient::builder("my.prefix", NopMetricSink)
    .with_tag("env", "prod")
    .with_tag("app", "auth")
    .build();

let res = client.count_with_tags("my.counter", 29)
    .with_tag("host", "web03.example.com")
    .with_tag_value("beta-test")
    .try_send();

assert_eq!(
    concat!(
        "my.prefix.my.counter:29|c|#",
        "env:prod,",
        "app:auth,",
        "host:web03.example.com,",
        "beta-test"
    ),
    res.unwrap().as_metric_str()
);

Value Packing

Value packing allows multiple values to be sent as a single metric for histograms, distributions, and timer types. The Cadence client accepts Vec<T> for histogram, distribution, and timer methods and will format multiple values as described below. Note that this feature is a Datadog extension and so may not be supported by your server. It is supported by versions >=v6.25.0 && <v7.0.0 or >=v7.25.0 of the Datadog agent.

Packed metrics have the following format:

<METRIC_NAME>:<VALUE1>:<VALUE2>:<VALUE3>|<TYPE>|#<TAG_KEY_1>:<TAG_VALUE_1>,<TAG_2>`

See the Datadog Docs for more information.

use cadence::prelude::*;
use cadence::{Metric, StatsdClient, NopMetricSink};

let client = StatsdClient::from_sink("my.prefix", NopMetricSink);

let res = client.distribution_with_tags("my.distribution", vec![29, 30, 31, 32])
    .with_tag("host", "web03.example.com")
    .with_tag_value("beta-test")
    .try_send();

assert_eq!(
    concat!(
        "my.prefix.my.distribution:29:30:31:32|d|#",
        "host:web03.example.com,",
        "beta-test"
    ),
    res.unwrap().as_metric_str()
);

Implemented Traits

Each of the methods that the Cadence StatsdClient struct uses to send metrics are implemented as a trait. There is also a trait that combines all of these other traits. If we want, we can just use one of the trait types to refer to the client instance. This might be useful to you if you'd like to swap out the actual Cadence client with a dummy version when you are unit testing your code or want to abstract away all the implementation details of the client being used behind a trait and pointer.

Each of these traits are exported in the prelude module. They are also available in the main module but aren't typically used like that.

use cadence::prelude::*;
use cadence::{StatsdClient, UdpMetricSink, DEFAULT_PORT};

pub struct User {
    id: u64,
    username: String,
    email: String
}

// Here's a simple DAO (Data Access Object) that doesn't do anything but
// uses a metric client to keep track of the number of times the
// 'getUserById' method gets called.
pub struct MyUserDao {
    metrics: Box<dyn MetricClient>
}

impl MyUserDao {
    // Create a new instance that will use the StatsdClient
    pub fn new<T: MetricClient + 'static>(metrics: T) -> MyUserDao {
        MyUserDao { metrics: Box::new(metrics) }
    }

    /// Get a new user by their ID
    pub fn get_user_by_id(&self, id: u64) -> Option<User> {
        self.metrics.incr("getUserById");
        None
    }
}

// Create a new Statsd client that writes to "metrics.example.com"
let host = ("metrics.example.com", DEFAULT_PORT);
let socket = UdpSocket::bind("0.0.0.0:0").unwrap();
let sink = UdpMetricSink::from(host, socket).unwrap();
let metrics = StatsdClient::from_sink("counter.example", sink);

// Create a new instance of the DAO that will use the client
let dao = MyUserDao::new(metrics);

// Try to lookup a user by ID!
match dao.get_user_by_id(123) {
    Some(u) => println!("Found a user!"),
    None => println!("No user!")
};

Quiet Metric Sending and Error Handling

When sending metrics sometimes you don't really care about the Result of trying to send it or maybe you just don't want to deal with it inline with the rest of your code. In order to handle this, Cadence allows you to set a default error handler. This handler is invoked when there are errors sending metrics so that the calling code doesn't have to deal with them.

An example of configuring an error handler and an example of when it might be invoked is given below.

use cadence::prelude::*;
use cadence::{MetricError, StatsdClient, NopMetricSink};

fn my_error_handler(err: MetricError) {
    println!("Metric error! {}", err);
}

let client = StatsdClient::builder("prefix", NopMetricSink)
    .with_error_handler(my_error_handler)
    .build();

// When sending metrics via the `MetricBuilder` used for assembling tags,
// callers may opt into sending metrics quietly via the `.send()` method
// as opposed to the `.try_send()` method
client.count_with_tags("some.counter", 42)
    .with_tag("region", "us-east-2")
    .send();

Custom Metric Sinks

The Cadence StatsdClient uses implementations of the MetricSink trait to send metrics to a metric server. Most users of the Cadence library probably want to use the QueuingMetricSink wrapping an instance of the BufferedMetricSink.

However, maybe you want to do something not covered by an existing sink. An example of creating a custom sink is below.

use std::io;
use cadence::prelude::*;
use cadence::{StatsdClient, MetricSink, DEFAULT_PORT};

pub struct MyMetricSink;

impl MetricSink for MyMetricSink {
    fn emit(&self, metric: &str) -> io::Result<usize> {
        // Your custom metric sink implementation goes here!
        Ok(0)
    }
}

let sink = MyMetricSink;
let client = StatsdClient::from_sink("my.prefix", sink);

client.count("my.counter.thing", 42);
client.time("my.method.time", 25);
client.incr("some.other.counter");

Custom UDP Socket

Most users of the Cadence StatsdClient will be using it to send metrics over a UDP socket. If you need to customize the socket, for example you want to use the socket in blocking mode but set a write timeout, you can do that as demonstrated below.

use std::net::UdpSocket;
use std::time::Duration;
use cadence::prelude::*;
use cadence::{StatsdClient, UdpMetricSink, DEFAULT_PORT};

let socket = UdpSocket::bind("0.0.0.0:0").unwrap();
socket.set_write_timeout(Some(Duration::from_millis(1))).unwrap();

let host = ("metrics.example.com", DEFAULT_PORT);
let sink = UdpMetricSink::from(host, socket).unwrap();
let client = StatsdClient::from_sink("my.prefix", sink);

client.count("my.counter.thing", 29);
client.time("my.service.call", 214);
client.incr("some.event");
client.set("users.uniques", 42);

Unix Sockets

Cadence also supports using Unix datagram sockets with the UnixMetricSink or BufferedUnixMetricSink. Unix sockets can be used for sending metrics to a server or agent running on the same machine (physical machine, VM, containers in a pod) as your application. Unix sockets are somewhat similar to UDP sockets with a few important differences:

• Sending metrics on a socket that doesn't exist or is not being listened to will result in an error.
• Metrics sent on a connected socket are guaranteed to be delievered (i.e. they are reliable as opposed to UDP sockets). However, it's still possible that the metrics won't be read by the server due to a variety of environment and server specific reasons.

An example of using the sinks is given below.

use std::os::unix::net::UnixDatagram;
use cadence::prelude::*;
use cadence::{StatsdClient, BufferedUnixMetricSink};

let socket = UnixDatagram::unbound().unwrap();
socket.set_nonblocking(true).unwrap();
let sink = BufferedUnixMetricSink::from("/run/statsd.sock", socket);
let client = StatsdClient::from_sink("my.prefix", sink);

client.count("my.counter.thing", 29);
client.time("my.service.call", 214);
client.incr("some.event");
client.set("users.uniques", 42);

NOTE: This feature is only available on Unix platforms (Linux, BSD, MacOS).

Other

For more information about Cadence, see the README in the repository root.