Menu
Documentationbreadcrumb arrow Grafana Mimirbreadcrumb arrow Sendbreadcrumb arrow Native histograms
Open source RSS

Send native histograms to Mimir

Prometheus native histograms is a data type in the Prometheus ecosystem that makes it possible to produce, store, and query a high-resolution histogram of observations.

Native histograms are different from classic Prometheus histograms in a number of ways:

  • Native histogram bucket boundaries are calculated by a formula that depends on the scale (resolution) of the native histogram, and are not user defined. The calculation produces exponentially increasing bucket boundaries. For details, refer to Bucket boundary calculation.
  • Native histogram bucket boundaries might change (widen) dynamically if the observations result in too many buckets. For details, refer to Limit the number of buckets.
  • Native histogram bucket counters only count observations inside the bucket boundaries, whereas the classic histogram buckets only have an upper bound called le and count all observations in the bucket and all lower buckets (cumulative).
  • An instance of a native histogram metric only requires a single time series, because the buckets, sum of observations, and the count of observations are stored in a single data type called native histogram rather than in separate time series using the float data type. Thus, there are no <metric>_bucket, <metric>_sum, and <metric>_count series. There is only <metric> time series.
  • Querying native histograms via the Prometheus query language (PromQL) uses a different syntax. For details, refer to Visualize native histograms and functions.

For an introduction to native histograms, watch the Native Histograms in Prometheus presentation.

Advantages and disadvantages

There are advantages and disadvantages of using native histograms compared to the classic Prometheus histograms. For more information and a real example, refer to the Prometheus Native Histograms in Production video.

Advantages

  • Simpler instrumentation: you do not need to think about bucket boundaries because they are created automatically.
  • Better resolution in practice: custom bucket layouts are usually not high resolution.
  • Native histograms are compatible with each other: they have an automatic layout, which makes them easy to combine.

    Note

    The operation might scale down an operand to lower resolution to match the other operand.

Disadvantages

  • Observations might be distributed in a way that is not a good fit for the exponential bucket schema, such as sound pressure measured in decibels, which are already logarithmic.
  • If converting from an externally represented histogram with specific bucket boundaries, there is generally no precise match with the bucket boundaries of the native histogram, and in which case you need to use interpolation.
  • There is no way to set an arbitrary bucket boundary, such as one that is particularly interesting for an SLO definition. Generally, ratios of observations above or below a given threshold have to be estimated by interpolation, rather than being precise in the case for a classic histogram with a configured bucket boundary at a given threshold.

The preceding problems are mitigated by high resolution, which native histograms can provide at a much lower resource cost compared to classic histograms.

Instrument application with Prometheus client libraries

The following examples have some reasonable defaults to define a new native histogram metric. The examples use the Go client library version 1.16 and the Java client library 1.0.

Note

Native histogram options can be added to existing classic histograms to get both the classic and native histogram at the same time. Refer to Migrate from classic histograms.
Go
histogram := prometheus.NewHistogram(
   prometheus.HistogramOpts{
      Name: "request_latency_seconds",
      Help: "Histogram of request latency in seconds",
      NativeHistogramBucketFactor: 1.1,
      NativeHistogramMaxBucketNumber: 100,
      NativeHistogramMinResetDuration: 1*time.Hour,
})
java
static final Histogram requestLatency = Histogram.build()
     .name("requests_latency_seconds")
     .help("Histogram of request latency in seconds")
     .nativeOnly()
     .nativeInitialSchema(3)
     .nativeMaxNumberOfBuckets(100)
     .nativeResetDuration(1, TimeUnit.HOURS)
     .register();

In Go, the NativeHistogramBucketFactor option sets an upper limit of the relative growth from one bucket to the next. The value 1.1 means that a bucket is at most 10% wider than the next smaller bucket. The currently supported values range from 1.0027 or 0.27% up to 65536 or 655%. For more detailed explanation, refer to Bucket boundary calculation.

Some of the resulting buckets for factor 1.1 rounded to two decimal places are:

…, (0.84, 0.92], (0.92, 1], (1, 1.09], (1.09, 1.19], (1.19, 1.30], …

…, (76.1, 83], (83, 91], (91, 99], …

…, (512, 558], (558, 608], (608, 663], …

In Java .nativeInitialSchema using schema value 3 results in the same bucket boundaries. For more information about the schema supported in Java, consult the documentation for nativeInitialSchema.

The value of NativeHistogramMaxBucketNumber/nativeMaxNumberOfBuckets limits the number of buckets produced by the observations. This can be especially useful if the receiver side is limiting the number of buckets that can be sent. For more information about the bucket limit refer to Limit the number of buckets.

The duration in NativeHistogramMinResetDuration/nativeResetDuration will prohibit automatic counter resets inside that period. Counter resets are related to the bucket limit, for more information refer to Limit the number of buckets.

Scrape and send native histograms with Prometheus

Use the latest version of Prometheus or at least version 2.47.

  1. To enable scraping native histograms from the application, you need to enable native histograms feature via a feature flag on the command line:

    bash
    prometheus --enable-feature=native-histograms
  2. The above flag will make Prometheus detect and scrape native histograms, but ignores classic histogram version of those metrics that have native histogram defined as well. Classic histograms without native histogram definitions are not effected. To keep scraping the classic histogram version of native histogram metrics you need to set scrape_classic_histograms to true in your scrape jobs, for example:

    yaml
    scrape_configs:
      - job_name: myapp
        scrape_classic_histograms: true

    in your scrape jobs, to get both histogram version.

    Note

    Native histograms don’t have a textual presentation at the moment on the application’s /metrics endpoint, thus Prometheus negotiates a Protobuf protocol transfer in this case.

    Note

    In certain situations, the protobuf parsing changes the number formatting of the le labels of conventional histograms and the quantile labels of summaries. Typically, this happens if the scraped target is instrumented with client_golang, provided that promhttp.HandlerOpts.EnableOpenMetrics is set to false. In such cases, integer label values are represented as quantile="1" or le="2" omitting the zero fractional. However, the protobuf parsing changes the representation to always include a fractional (following the OpenMetrics specification), so the examples above become quantile="1.0" and le="2.0" after ingestion into Prometheus, which changes the identity of the metric from what was originally ingested.

    For more information, refer to Feature Flags Native Histograms in the Prometheus documentation.

  3. To be able to send native histograms to a Prometheus remote write compatible receiver, for example Grafana Cloud Metrics, Mimir, etc, set send_native_histograms to true in the remote write configuration, for example:

    yaml
    remote_write:
      - url: http://.../api/prom/push
        send_native_histograms: true

Scrape and send native histograms with Grafana Alloy

Use the latest version of Grafana Alloy.

  1. To scrape native histograms, you need to set the scrape_protocols argument in the prometheus.scrape component to specify PrometheusProto as the first protocol to negotiate and scrape_classic_histograms = true in order to scrape both classic and native histograms.

     scrape_protocols = ["PrometheusProto", "OpenMetricsText1.0.0", "OpenMetricsText0.0.1", "PrometheusText0.0.4"]
     scrape_classic_histograms = true

    For more information, refer to prometheus.scrape in the Grafana Alloy documentation.

    Note

    In certain situations, the protobuf parsing changes the number formatting of the le labels of conventional histograms and the quantile labels of summaries. Typically, this happens if the scraped target is instrumented with client_golang, provided that promhttp.HandlerOpts.EnableOpenMetrics is set to false. In such cases, integer label values are represented as quantile="1" or le="2" omitting the zero fractional. However, the protobuf parsing changes the representation to always include a fractional (following the OpenMetrics specification), so the examples above become quantile="1.0" and le="2.0" after ingestion into Prometheus, which changes the identity of the metric from what was originally ingested.

    For more information, refer to Feature Flags Native Histograms in the Prometheus documentation.

  2. To send native histograms to a Prometheus remote write compatible receiver, such as Grafana Cloud Metrics or Mimir, set the send_native_histograms argument to true in the prometheus.remote_write component. For example:

    prometheus.remote_write "mimir" {
      endpoint {
        url = "http://.../api/prom/push"
        send_native_histograms = true
      }
    }

Migrate from classic histograms

To ease the migration process, you can keep the custom bucket definition of a classic histogram and add native histogram buckets at the same time.

  1. Add the native histogram definition to an existing histogram in the instrumentation.

  2. If the existing histogram doesn’t have buckets defined, add the default buckets to keep the classic histogram.

    Code examples with both classic and native histogram defined for the same metric:

    Go
    histogram := prometheus.NewHistogram(
       prometheus.HistogramOpts{
           Name: "request_latency_seconds",
           Help: "Histogram of request latency in seconds",
           Buckets: prometheus.DefBuckets,  // If buckets weren't already defined.
           NativeHistogramBucketFactor: 1.1,
           NativeHistogramMaxBucketNumber: 100,
           NativeHistogramMinResetDuration: 1*time.Hour,
    })
    java
    static final Histogram requestLatency = Histogram.build()
       .name("requests_latency_seconds")
       .help("Histogram of request latency in seconds")
       .classicUpperBounds(Histogram.Builder.DEFAULT_CLASSIC_UPPER_BOUNDS)  // If upper bounds weren't already defined.
       .nativeInitialSchema(3)
       .nativeMaxNumberOfBuckets(100)
       .nativeResetDuration(1, TimeUnit.HOURS)
       .register();
  3. Let Prometheus or Grafana Alloy scrape both classic and native histograms for metrics that have both defined.

  4. Send native histograms to remote write, along with the existing classic histograms.

  5. Modify dashboards to use the native histogram metrics. Refer to Visualize native histograms for more information.

    Use one of the following strategies to update dashbaords.

    • (Recommended) Add new dashboards with the new native histogram queries. This solution requires looking at different dashboards for data before and after the migration, until data before the migration is removed due to passing its retention time. You can publish the new dashboard when sufficient time has passed to serve users with the new data.

    • Add a dashboard variable to your dashboard to enable switching between classic histograms and native histograms. There isn’t support for selectively enabling and disabling queries in Grafana (issue 79848). As a workaround, add the dashboard variable latency_metrics, for example, and assign it a value of either -1 or 1. Then, add the following two queries to the panel:

      <classic_query> < ($latency_metrics * +Inf)
      <native_query> < ($latency_metrics * -Inf)

      Where classic_query is the original query and native_query is the same query using native histogram query syntax. This technique is employed in Mimir’s dashboards. For an example, refer to the Overview dashboard in the Mimir repository.

      This solution allows users to switch between the classic histogram and the native histogram without going to a different dashboard.

    • Replace the existing classic queries with modified queries. For example, replace:

      <classic_query>

      with

      <native_query> or <classic_query>

      Where classic_query is the original query and native_query is the same query using native histogram query syntax.

      Warning

      Using the PromQL operator or can lead to unexpected results. For example, if a query uses a range of seven days, such as sum(rate(http_request_duration_seconds[7d])), then this query returns a value as soon as there are two native histograms samples present before the end time specified in the query. In this case, the seven day rate is calculated from a couple of minutes, rather than seven days, worth of data. This results in an inaccuracy in the graph around the time you started scraping native histograms.
  6. Start adding new recording rules and alerts to use native histograms. Do not remove the old recording rules and alerts at this time.

  7. It is important to keep scraping both classic and native histograms for at least the period of the longest range in your recording rules and alerts, plus one day. This is the minimum amount of time, but it’s recommended to keep scraping both data types until the new rules and alerts can be verified.

    For example, if you have an alert that calculates the rate of requests, such as sum(rate(http_request_duration_seconds[7d])), this query looks at the data from the last seven days plus the Prometheus lookback period. When you start sending native histograms, the data isn’t there for the entire seven days, and therefore, the results might be unreliable for alerting.

  8. After configuring native histogram collection, choose one of the following ways to stop collecting classic histograms.

  9. Clean up recording rules and alerts by deleting the classic histogram version of the rule or alert.

Bucket boundary calculation

This section assumes that you are familiar with basic algebra. Native histogram bucket boundaries are calculated from an exponential formula with a base of 2.

Native histogram samples have three different kind of buckets, for any observed value the value is counted towards one kind of bucket.

  • A zero bucket, which contains the count of observations whose absolute value is smaller or equal to the zero threshold.

Zero threshold definition

  • Positive buckets, which contain the count of observations with a positive value that is greater than the lower bound and less or equal to the upper bound of a bucket.

Positive bucket definition

where the index can be a positive or negative integer resulting in boundaries above 1 and fractions below 1. The schema either directly specified out of [-4, -3, -2, -1, 0, 1, 2, 3, 4, 5, 6, 7, 8] at instrumentation time or it is the largest number chosen from the list in such way that

Factor equation

for example for factor 1.1:

Factor 1.1 equation

Table of schema to factor:

schemafactorschemafactor
-46553631.0905
-325641.0443
-21651.0219
-1461.0109
0271.0054
11.414281.0027
21.1892
  • Negative buckets, which contain the count of observations with a negative value that is smaller than the upper bound and greater than or equal to the lower bound of a bucket.

Negative bucket definition

where the schema is chosen as above.

Limit the number of buckets

The server scraping or receiving native histograms over remote write may limit the number of native histogram buckets it accepts. The server may reject or downscale (reduce resolution and merge adjacent buckets). Even if that wasn’t the case, storing and emitting potentially unlimited number of buckets isn’t practical.

The instrumentation libraries of Prometheus have automation to keep the number of buckets down, provided that the maximum bucket number option is used, such as NativeHistogramMaxBucketNumber in Go.

After the set maximum is exceeded, the following strategy is enacted:

  1. First, if the last reset (or the creation) of the histogram is at least the minimum reset duration ago, then the whole histogram is reset to its initial state (including classic buckets). This only works if the minimum reset duration was set (NativeHistogramMinResetDuration in Go).

  2. If less time has passed, or if the minimum reset duration is zero, no reset is performed. Instead, the zero threshold is increased sufficiently to reduce the number of buckets to or below the maximum bucket number, but not to more than the maximum zero threshold (NativeHistogramMaxZeroThreshold in Go). Thus, if the threshold is at or above the maximum threshold already nothing happens at this step.

  3. After that, if the number of buckets still exceeds maximum bucket number, the resolution of the histogram is reduced by doubling the width of all the buckets (up to a growth factor between one bucket to the next of 2^(2^4) = 65536, refer to Bucket boundary calculation).

  4. Any increased zero threshold or reduced resolution is reset back to their original values once the minimum reset duration has passed (since the last reset or the creation of the histogram).