2 unstable releases
0.2.0 | Aug 7, 2024 |
---|---|
0.1.0 | Jul 5, 2024 |
#67 in Development tools
41,731 downloads per month
Used in 71 crates
(38 directly)
130KB
2.5K
SLoC
Vise – Typesafe Metrics Client
This library provides a wrapper for defining and reporting metrics in Rust libraries and applications.
It is based on the prometheus-client
library, augmenting it with higher-level / more eloquent functionality.
Features
- Allows registering and reporting metrics in an idiomatic and typesafe manner.
- Allows testing metrics by accessing their current values. (Note: Accessing metric data is not implemented for histograms yet.)
What are metrics, anyway
Metrics are numerical measurements taken over time. Metrics are defined and collected in an application and reported to an external system, Prometheus, from which they can be accessed using e.g. Grafana dashboards.
Prometheus and compatible systems supports 3 main metric types:
- Counters are monotonically increasing integer values
- Gauges are integer or floating-point values that can go up or down. Logically, a reported gauge value can be treated as valid until the next value is reported.
- Histograms are floating-point values counted in configurable buckets. Logically, a histogram observes a certain probability distribution, and observations are transient (unlike gauge values).
Metrics of all types can be supplied with labels. Each set of labels defines a separate metric. Thus, label space should be reasonably small.
Besides these main types, Prometheus and this library support an additional info metric type. It should be used to observe values that do not change during the program lifetime (component configurations, metadata like app version / git commit, etc.). Values for this metric type are encoded as labels. Conceptually, an info metric is similar to a gauge with a constant value 1.
Usage
Add this to your Crate.toml:
[dependencies]
vise = "0.2.0"
Defining and reporting metrics
Metrics are defined as structs, with each field corresponding to a metric or a family of metrics:
use vise::*;
use std::{fmt, time::Duration};
/// Metrics defined by the library or application. A single app / lib can define
/// multiple metric structs.
#[derive(Debug, Metrics)]
#[metrics(prefix = "my_app")]
// ^ Prefix added to all field names to get the final metric name (e.g., `my_app_latencies`).
pub(crate) struct MyMetrics {
/// Simple counter. Doc comments for the fields will be reported
/// as Prometheus metric descriptions.
pub counter: Counter,
/// Integer-valued gauge. Unit will be reported to Prometheus and will influence metric name
/// by adding the corresponding suffix to it (in this case, `_bytes`).
#[metrics(unit = Unit::Bytes)]
pub gauge: Gauge<u64>,
/// Group of histograms with the "method" label.
/// Each `Histogram` or `Family` of `Histogram`s must define buckets; in this case,
/// we use default buckets for latencies.
#[metrics(buckets = Buckets::LATENCIES, labels = ["method"])]
pub latencies: LabeledFamily<&'static str, Histogram<Duration>>,
}
// Commonly, it makes sense to make metrics available using a static:
#[vise::register]
static MY_METRICS: Global<MyMetrics> = Global::new();
// Metrics are singletons globally available using the `instance()` method.
MY_METRICS.counter.inc();
assert_eq!(MY_METRICS.counter.get(), 1); // Useful for testing
let latency = MY_METRICS.latencies[&"test"].start();
// Do some work...
let latency: Duration = latency.observe();
// `latency` can be used in logging etc.
See crate docs for more examples.
Testing metrics
Depending on how you report metrics (e.g., whether the global state is used), testing metrics may require refactoring.
- You may pass around references to the metric type(s) in the logic under test so that these types can be injected and then checked by tests.
- Alternatively, your logic may produce statistics that are then reported as metrics (this may be beneficial for performance as well). In this case, produced statistics can be checked by tests.
Best practices
See also: Prometheus guidelines
- Metrics and metric labels should be named in snake_case. (This should be enforced by Clippy and checks performed in
the
Metrics
derive macro.) - Metrics should start with a prefix or a sequence of prefixes describing the domain / subdomains owning the metric.
Prefixes should be separated by a single
_
char. - Metrics with a unit should have a corresponding suffix (e.g.,
_seconds
). This suffix is automatically added to the metric name if you specify its unit; you must not specify it manually. - Label names should not repeat the metric name.
- Label values for each label should have reasonably low cardinality.
- If a label value encodes to a string (as opposed to an integer, integer range etc.), it should use snake_case.
- Metrics in a
Family
should have uniform meaning. If aFamily
can be documented without going into label specifics, you're usually on a right track.
Example: RocksDB size metrics
Suppose we want to report live and total data sizes for RocksDB instances that live in our application. We may want to define:
- Families of gauges (since data sizes logically persist until the next size is reported)
- ...with
rocksdb_
prefix - ...separate families for live and total data sizes (since they measure 2 distinct things)
- ...with
db
andcf
labels specifying the database ID and column family name (the database ID should be globally unique; column families will probably differ amongdb
values) - ...with
Unit::Bytes
(since data sizes are measured in bytes)
Thus, we might have the following metrics:
rocksdb_live_data_size_bytes{db="merkle_tree",cf="default"} 123456789
rocksdb_live_data_size_bytes{db="merkle_tree",cf="stale_keys"} 123456
rocksdb_total_data_size_bytes{db="merkle_tree",cf="default"} 130000000
rocksdb_total_data_size_bytes{db="merkle_tree",cf="stale_keys"} 130000
License
Distributed under the terms of either
- Apache License, Version 2.0, (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
- MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)
at your option.
Dependencies
~1–6MB
~35K SLoC