Crate hotmic

High-speed metrics collection library.

hotmic provides a generalized metrics collection library targeted at users who want to log metrics at high volume and high speed.

Design

The library follows a pattern of "senders" and a "receiver."

Callers create a Receiver, which acts as a contained unit: metric registration, aggregation, and summarization. The Receiver is intended to be spawned onto a dedicated background thread.

From a Receiver, callers can create a Sink, which allows registering facets -- or interests -- in a given metric, along with sending the metrics themselves. All metrics need to be pre-registered, in essence, with the receiver, which allows us to know which aspects of a metric to track: count, value, or percentile.

A Sink can be cheaply cloned and does not require a mutable reference to send metrics, and so callers have great flexibility in being able to control their resource consumption when it comes to sinks. Receiver also allows configuring the capacity of the underlying channels to finely tune resource consumption.

Being based on [crossbeam-channel] allows us to process close to five million metrics per second on a single core, with very low ingest latencies: 325-350ns on average at full throughput.

Metrics

hotmic supports counters, gauges, and histograms.

A counter is a single value that can be updated with deltas to increase or decrease the value. This would be your typical "messages sent" or "database queries executed" style of metric, where the value changes over time.

A gauge is also a single value but does not support delta updates. When a gauge is set, the value sent becomes the value of the gauge. Gauges can be useful for metrics that measure a point-in-time value, such as "connected clients" or "running queries". While those metrics could also be represented by a count, gauges can be simpler in cases where you're already computing and storing the value, and simply want to expose it in your metrics.

A histogram tracks the distribution of values: how many values were between 0-5, between 6-10, etc. This is the canonical way to measure latency: the time spent running a piece of code or servicing an operation. By keeping track of the individual measurements, we can better see how many are slow, fast, average, and in what proportions.

use hotmic::{Facet, Receiver};
use std::thread;
let receiver = Receiver::builder().build();
let sink = receiver.get_sink();

// We have to register the metrics we care about so that they're properly tracked!
sink.add_facet(Facet::Count("widget"));
sink.add_facet(Facet::Gauge("red_balloons"));
sink.add_facet(Facet::TimingPercentile("db.gizmo_query"));
sink.add_facet(Facet::Count("db.gizmo_query"));
sink.add_facet(Facet::ValuePercentile("buf_size"));

// We can send a simple count, which is a signed value, so the value we give is applied as a
// delta to the underlying counter.  After these sends, "widgets" would be 3.
assert!(sink.update_count("widgets", 5).is_ok());
assert!(sink.update_count("widgets", -3).is_ok());
assert!(sink.update_count("widgets", 1).is_ok());

// We can update a gauge.  This is just a point-in-time value so the last "write" of this
// metric is what the value will be, and it will stay at that value until changed.
assert!(sink.update_value("red_balloons", 99).is_ok());

// We can update a timing percentile.  For timing, you also must measure the start and end
// time using the built-in `Clock` exposed by the sink.  The receiver internally converts the
// raw values to calculate the actual wall clock time (in nanoseconds) on your behalf, so you
// can't just pass in any old number.. otherwise you'll get erroneous measurements!
let start = sink.clock().start();
thread::sleep_ms(10);
let end = sink.clock().end();
let rows = 42;

// This would just set the timing:
assert!(sink.update_timing("db.gizmo_query", start, end).is_ok());

// This would set the timing and also let you provide a customized count value.  Being able to
// specify a count is handy when tracking things like the time it took to execute a database
// query, along with how many rows that query returned:
assert!(sink
    .update_timing_with_count("db.gizmo_query", start, end, rows)
    .is_ok());

// Finally, we can update a value percentile.  Technically speaking, value percentiles aren't
// fundamentally different from timing percentiles.  If you use a timing percentile, we do the
// math for you of getting the time difference, and we make sure the metric name has the right
// unit suffix so you can tell it's measuring time, but other than that, nearly identical!
let buf_size = 4096;
assert!(sink.update_value("buf_size", buf_size).is_ok());

Facets

Facets are the way callers specify what they're interested in. Without any other configuration, you could send any metric you want but nothing would happen; nothing would be recorded.

Facets correspond roughly to the metric types, with the exception of the difference between timing percentiles and value percentiles, which both are histogram-based but differ in how we render their metric labels.

Thus, if you want to record a counter, you would register a counter facet for the given metric key, and if you want to track latency for a given operation, you would register a timing percentile for the metric key used.

Facets and scoping (explained below) are intrinsically tied together, so facets need to be registered directly on the sink they'll be used from in order to ensure that the facet matches the scope of the sink:

use hotmic::{Facet, Receiver};
let receiver = Receiver::builder().build();

// This sink has no scope aka the root scope.  We can register facets on this sink without a
// problem, but if get a scoped sink from this one, and sent the same metric name, the scopes
// would not line up, and the metric wouldn't be registered for storage.
let root_sink = receiver.get_sink();
root_sink.add_facet(Facet::Count("widgets"));
assert!(root_sink.update_count("widgets", 42).is_ok());

// Make a new scoped sink.  If we tried to send to this new sink, without reregistering our
// facets, our metrics wouldn't be stored at all.
let scoped_sink = root_sink.scoped("party").unwrap();

// Register the facet, and we're all good.
scoped_sink.add_facet(Facet::Count("widgets"));
assert!(scoped_sink.update_count("widgets", 43).is_ok());

Scopes

Metrics can be scoped, not unlike loggers, at the Sink level. This allows sinks to easily nest themselves without callers ever needing to care about where they're located.

This feature is a simpler approach to tagging: while not as semantically rich, it provides the level of detail necessary to distinguish a single metric between multiple callsites.

For example, after getting a Sink from the Receiver, we can easily nest ourselves under the root scope and then send some metrics:

use hotmic::Receiver;
let receiver = Receiver::builder().build();

// This sink has no scope aka the root scope.  The metric will just end up as "widgets".
let root_sink = receiver.get_sink();
assert!(root_sink.update_count("widgets", 42).is_ok());

// This sink is under the "secret" scope.  Since we derived ourselves from the root scope,
// we're not nested under anything, but our metric name will end up being "secret.widgets".
let scoped_sink = root_sink.scoped("secret").unwrap();
assert!(scoped_sink.update_count("widgets", 42).is_ok());

// This sink is under the "supersecret" scope, but we're also nested!  The metric name for this
// sample will end up being "secret.supersecret.widget".
let scoped_sink_two = scoped_sink.scoped("supersecret").unwrap();
assert!(scoped_sink_two.update_count("widgets", 42).is_ok());

// Sinks retain their scope even when cloned, so the metric name will be the same as above.
let cloned_sink = scoped_sink_two.clone();
assert!(cloned_sink.update_count("widgets", 42).is_ok());

Structs

Configuration	A configuration builder for Receiver.
Controller	Dedicated handle for performing operations on a running Receiver.
Percentile	A labeled percentile.
Receiver	Metrics receiver which aggregates and processes samples.
Sink	Handle for sending metric samples into the receiver.
Snapshot	A point-in-time view of metric data.

Enums

Facet	Type of computation against aggregated/processed samples.
SinkError	Erorrs during sink creation.