Crate hdrhistogram

source ·
Expand description

HdrSample is a port of Gil Tene’s HdrHistogram to native Rust. It provides recording and analyzing of sampled data value counts across a large, configurable value range with configurable precision within the range. The resulting “HDR” histogram allows for fast and accurate analysis of the extreme ranges of data with non-normal distributions, like latency.

HdrHistogram

What follows is a description from the HdrHistogram website. Users are encouraged to read the documentation from the original Java implementation, as most of the concepts translate directly to the Rust port.

HdrHistogram supports the recording and analyzing of sampled data value counts across a configurable integer value range with configurable value precision within the range. Value precision is expressed as the number of significant digits in the value recording, and provides control over value quantization behavior across the value range and the subsequent value resolution at any given level.

For example, a Histogram could be configured to track the counts of observed integer values between 0 and 3,600,000,000 while maintaining a value precision of 3 significant digits across that range. Value quantization within the range will thus be no larger than 1/1,000th (or 0.1%) of any value. This example Histogram could be used to track and analyze the counts of observed response times ranging between 1 microsecond and 1 hour in magnitude, while maintaining a value resolution of 1 microsecond up to 1 millisecond, a resolution of 1 millisecond (or better) up to one second, and a resolution of 1 second (or better) up to 1,000 seconds. At it’s maximum tracked value (1 hour), it would still maintain a resolution of 3.6 seconds (or better).

HDR Histogram is designed for recording histograms of value measurements in latency and performance sensitive applications. Measurements show value recording times as low as 3-6 nanoseconds on modern (circa 2014) Intel CPUs. The HDR Histogram maintains a fixed cost in both space and time. A Histogram’s memory footprint is constant, with no allocation operations involved in recording data values or in iterating through them. The memory footprint is fixed regardless of the number of data value samples recorded, and depends solely on the dynamic range and precision chosen. The amount of work involved in recording a sample is constant, and directly computes storage index locations such that no iteration or searching is ever involved in recording data values.

If you are looking for FFI bindings to HdrHistogram_c, you want the hdrhistogram_c crate instead.

Interacting with the library

HdrSample’s API follows that of the original HdrHistogram Java implementation, with some modifications to make its use more idiomatic in Rust. The description in this section has been adapted from that given by the Python port, as it gives a nicer first-time introduction to the use of HdrHistogram than the Java docs do.

HdrSample is generally used in one of two modes: recording samples, or querying for analytics. In distributed deployments, the recording may be performed remotely (and possibly in multiple locations), to then be aggregated later in a central location for analysis.

Recording samples

A histogram instance is created using the ::new methods on the Histogram struct. These come in three variants: new, new_with_max, and new_with_bounds. The first of these only sets the required precision of the sampled data, but leaves the value range open such that any value may be recorded. A Histogram created this way (or one where auto-resize has been explicitly enabled) will automatically resize itself if a value that is too large to fit in the current dataset is encountered. new_with_max sets an upper bound on the values to be recorded, and disables auto-resizing, thus preventing any re-allocation during recording. If the application attempts to record a larger value than this maximum bound, the record call will return an error. Finally, new_with_bounds restricts the lowest representable value of the dataset, such that a smaller range needs to be covered (thus reducing the overall allocation size).

For example the example below shows how to create a Histogram that can count values in the [1..3600000] range with 1% precision, which could be used to track latencies in the range [1 msec..1 hour]).

use hdrhistogram::Histogram;
let mut hist = Histogram::<u64>::new_with_bounds(1, 60 * 60 * 1000, 2).unwrap();

// samples can be recorded using .record, which will error if the value is too small or large
hist.record(54321).expect("value 54321 should be in range");

// for ergonomics, samples can also be recorded with +=
// this call will panic if the value is out of range!
hist += 54321;

// if the code that generates the values is subject to Coordinated Omission,
// the self-correcting record method should be used instead.
// for example, if the expected sampling interval is 10 msec:
hist.record_correct(54321, 10).expect("value 54321 should be in range");

Note the u64 type. This type can be changed to reduce the storage overhead for all the histogram bins, at the cost of a risk of saturating if a large number of samples end up in the same bin.

Querying samples

At any time, the histogram can be queried to return interesting statistical measurements, such as the total number of recorded samples, or the value at a given quantile:

use hdrhistogram::Histogram;
let hist = Histogram::<u64>::new(2).unwrap();
// ...
println!("# of samples: {}", hist.len());
println!("99.9'th percentile: {}", hist.value_at_quantile(0.999));

Several useful iterators are also provided for quickly getting an overview of the dataset. The simplest one is iter_recorded(), which yields one item for every non-empty sample bin. All the HdrHistogram iterators are supported in HdrSample, so look for the *Iterator classes in the Java documentation.

use hdrhistogram::Histogram;
let hist = Histogram::<u64>::new(2).unwrap();
// ...
for v in hist.iter_recorded() {
    println!("{}'th percentile of data is {} with {} samples",
        v.percentile(), v.value_iterated_to(), v.count_at_value());
}

Panics and error handling

As long as you’re using safe, non-panicking functions (see below), this library should never panic. Any panics you encounter are a bug; please file them in the issue tracker.

A few functions have their functionality exposed via AddAssign and SubAssign implementations. These alternate forms are equivalent to simply calling unwrap() on the normal functions, so the normal rules of unwrap() apply: view with suspicion when used in production code, etc.

Returns ResultPanics on errorFunctionality
h.record(v)h += vIncrement count for value v
h.add(h2)h += h2Add h2’s counts to h
h.subtract(h2)h -= h2Subtract h2’s counts from h

Other than the panicking forms of the above functions, everything will return Result or Option if it can fail.

usize limitations

Depending on the configured number of significant digits and maximum value, a histogram’s internal storage may have hundreds of thousands of cells. Systems with a 16-bit usize cannot represent pointer offsets that large, so relevant operations (creation, deserialization, etc) will fail with a suitable error (e.g. CreationError::UsizeTypeTooSmall). If you are using such a system and hitting these errors, reducing the number of significant digits will greatly reduce memory consumption (and therefore the need for large usize values). Lowering the max value may also help as long as resizing is disabled.

32- and above systems will not have any such issues, as all possible histograms fit within a 32-bit index.

Floating point accuracy

Some calculations inherently involve floating point values, like value_at_quantile, and are therefore subject to the precision limits of IEEE754 floating point calculations. The user- visible consequence of this is that in certain corner cases, you might end up with a bucket (and therefore value) that is higher or lower than it would be if the calculation had been done with arbitrary-precision arithmetic. However, double-precision IEEE754 (i.e. f64) is very good at its job, so these cases should be rare. Also, we haven’t seen a case that was off by more than one bucket.

To minimize FP precision losses, we favor working with quantiles rather than percentiles. A quantile represents a portion of a set with a number in [0, 1]. A percentile is the same concept, except it uses the range [0, 100]. Working just with quantiles means we can skip an FP operation in a few places, and therefore avoid opportunities for precision loss to creep in.

Limitations and Caveats

As with all the other HdrHistogram ports, the latest features and bug fixes from the upstream HdrHistogram implementations may not be available in this port. A number of features have also not (yet) been implemented:

  • Concurrency support (AtomicHistogram, ConcurrentHistogram, …).
  • DoubleHistogram.
  • The Recorder feature of HdrHistogram.
  • Value shifting (“normalization”).
  • Textual output methods. These seem almost orthogonal to HdrSample, though it might be convenient if we implemented some relevant traits (CSV, JSON, and possibly simple fmt::Display).

Most of these should be fairly straightforward to add, as the code aligns pretty well with the original Java/C# code. If you do decide to implement one and send a PR, please make sure you also port the test cases, and try to make sure you implement appropriate traits to make the use of the feature as ergonomic as possible.

Re-exports

Modules

  • Error types used throughout this library
  • Module containing the implementations of all Histogram iterators.
  • Serialization/deserialization support.
  • Synchronized types that allow access to a Histogram from multiple threads.

Structs

  • Histogram is the core data structure in HdrSample. It records values, and performs analytics.

Traits

  • This trait represents the operations a histogram must be able to perform on the underlying counter type. The ToPrimitive trait is needed to perform floating point operations on the counts (usually for quantiles). The FromPrimitive to convert back into an integer count. Partial ordering is used for threshholding, also usually in the context of quantiles.