prometheusreceiver

README

Prometheus Receiver Design Spec

Design Goal

Provide a seamless onboarding experience for users who are already familiar with Prometheus scrape config

Prometheus has a very powerful config system for user to config how Prometheus can scrape the metrics data from any application which expose a Prometheus format metrics endpoint. It provides very useful features like filtering unwanted metrics, relabeling tags..etc. The original Prometheus receiver of OpenTelemetry took the approach of using Prometheus' own scraper's source code as a library to achieve this goal. Overall the idea was great, however, the original implementation has a lot of glitches, it cannot be fixed by small patches. This new Prometheus receiver is going to follow the same idea of leveraging Prometheus sourcecode, with a proper implementation.

Map Prometheus metrics to the corresponding OpenTelemetry metrics properly

Prometheus receiver shall be able to map Prometheus metrics to OpenTelemetry's proto based metrics, it shall respect the original metric name, value, timestamp, as well as tags. It doesn't need to provide one-to-one mapping, since supported metric types are different from the two systems. However, it shall not drop data.

Parity between Prometheus and OpenTelemetry Prometheus exporter

Prometheus itself can also used as an exporter, that it can expose the metrics it scrape from other system with its own metrics endpoint, so is OpenTelemetry service. We shall be able to retain parity from the following two setups:

1. app -> prometheus -> metric-endpoint
2. app -> otelcol-with-prometheus-receiver -> otelcol-prometheus-exporter-metrics-endpoint

Prometheus Text Format Overview

Prometheus text format is a line orient format. For each non-empty line, which not begins with #, is a metric data point with includes a metric name and its value, which is of float64 type, as well as some optional data such as tags and timestamp, which is in milliseconds. For lines begin with #, they are either comments, which need to be filtered, or metadata, which including type hints and units that are usually indicating the beginning of a new individual metric or a group of new metrics. More details of Prometheus text format can be found from its official document.

Metric types

Based on this document, Prometheus supports the following 5 types of metrics:

• Counter
• Gauge
• Histogram
• Summary
• Untyped

However, this is not the whole story, from the implementation details of Prometheus scraper, which the receiver based on, it supports a couple more undocumented metrics types, including:

• Gaugehistogram
• Info
• Statset

More details can be found from the prometheus text parser source code

Metric Grouping

Other than metric types, the type hint comment and metric grouping are also important to know in order to parse Prometheus text metrics properly. From any Prometheus metrics endpoints, metrics are usually grouped together by starting with a comment section which includes some very important information like type hints about the metrics, and metrics points of the same group will have the same metric name but a different set of tag values, for example:

# HELP container_cpu_load_average_10s Value of container cpu load average over the last 10 seconds.
# TYPE container_cpu_load_average_10s gauge
container_cpu_load_average_10s{id="/",image="",name=""} 0
container_cpu_load_average_10s{id="/000-metadata",image="",name=""} 0
container_cpu_load_average_10s{id="/001-sysfs",image="",name=""} 0

The above example was taken from an cadvisor metric endpoint, the type hint tells that the name of this metric group is container_cpu_load_average_10s and it's of gague type. Then it follows by some individual metric points which are of the same metric name. For each individual metric within this group, they share the same set of tag keys, with unique value sets.

Prometheus Metric Scraper Anatomy

The metrics scraper is a component which is used to scrape remote Prometheus metric endpoints, it is also the component which Prometheus receiver is based on. It's important to understand how it works in order to implement the receiver properly.

Major components of Prometheus Scape package

• ScapeManager: the component which loads the scrape_config, and manage the scraping tasks

• ScrapePool: an object which manage scrapes for a sets of targets

• Scraper: a http client to fetch data from remote metrics endpoints

• Target: the remote metric endpoint, as well as related relabing settings and other metadata

• TextParser: a DFA style streaming decoder/parser for prometheus text format

• Appendable: it is used to acquire a storage appender instance at the beginning of each scrapeLoop run

• storage.Appender: an abstraction of the metric storage which can be a filesystem, a database or an remote endpoint...etc. As for OpenTelemetry prometheus receiver, this is also the interface we need to implement to provide a customized storage appender which is backed by metrics sink.

• ScrapeLoop: the actual scrape pipeline which performs the main scraping and ingestion logic.

Prometheus ScrapeLoop workflow explained

Each scraping cycle is trigger by an configured interval, its workflow is as shown in the flowchart below:

It basically does the following things in turn:

1. make a http call to fetch data from the binding target's metrics endpoint with scraper
2. acquired a storage appender instance with the Appendable interface
3. feed the data to a textParser
4. parse and feed metric data points to storage appender
5. commit if success or rollback
6. report task status

Implementing Prometheus storage.Appender with metrics sink

The storage.Appender interface

As discussed in the previous section, the storage.Appender is the most important piece of components for us to implement so as to bring the two worlds together. It has a very simple interface which is defined below:

type Appender interface {
  Add(l labels.Labels, t int64, v float64) (uint64, error)


  AddFast(l labels.Labels, ref uint64, t int64, v float64) error


  // Commit submits the collected samples and purges the batch.
  Commit() error


  Rollback() error
}

Note: the above code belongs to the Prometheus project, its license can be found here

One can see that the interface is very simple, it only has 4 methods: Add, AddFast, Commit and Rollback. The last two methods are easy to understand: Commit is called when the processing of the scraped page is completed and success, whereas Rollback is called if error occurs in between the process.

However for the two methods starting with 'Add', there's no document on the Prometheus project for how they should be used. By examining the scrapeLoop source code, as well as some storage.Appender implementations. It indicates that the first method Add is always used for the first time when a unique metrics, which means the combination of metric name and its tags are unique, is seen for the first time. The Add method can return a non zero reference number, then the scrapeLoop can cache this number with the metric's uniq signature. The next time, such as the next scrape cycle of the same target, when the metric is seen again by matching its signature, it will call the AddFast method with the cached reference number. This reference number might make sense to databases which has unique key as numbers, however, in our use case, it's not necessary, thus we can always return 0 ref number from the Add method to skip this caching mechanism.

Challenges and solutions

Even though the definition of this interface is very simple, however, to implement it properly is a bit challenging, given that every time the Add/AddFast method is called, it only provides the information about the current data point, the context of what metric group this data point belonging to is not provided, we have to keep track of it internally within the appender. And this is not the whole story, there are a couple other issues we need to address, including:

1. Have a way to link the Target with the current appender instance

The labels provided to the Add/AddFast methods dose not include some target specified information such as job name which is important construct the Node proto object of OpenTelemetry. The target object is not accessible from the Appender interface, however, we can get it from the ScrapeManager, when designing the appender, we need to have a way to inject the binding target into the appender instance.

1. Group metrics from the same family together

In OpenTelemetry, metric points of the same name are usually grouped together as one timeseries but different data points. It's important for the appender to keep track of the metric family changes, and group metrics of the same family together. Keep in mind that the Add/AddFast method is operated in a streaming manner, ScrapeLoop does not provide any direct hints on metric name change, the appender itself need to keep track of it. It's also important to know that for some special types such as histogram and summary, not all the data points have the same name, there are some special metric points has postfix like _sum and _count, we need to handle this properly, and do not consider this is a metric family change.

1. Group complex metrics such as histogram together in proper order

In Prometheus, a single aggregated type of metric data such as histogram and summary is represent by multiple metric data points, such as buckets and quantiles as well as the additional _sum and _count data. ScrapeLoop will feed them into the appender individually. The appender needs to have a way to bundle them together to transform them into a single Metric Datapoint Proto object.

1. Tags need to be handled carefully

ScrapeLoop strips out any tag with empty value, however, in OpenTelemetry, the tag keys is stored separately, we need to able to get all the possible tag keys of the same metric family before committing the metric family to the sink.

1. StartTimestamp and values of metrics of cumulative types

In OpenTelemetry, every metrics of cumulative type is required to have a StartTimestamp, which records when a metric is first recorded, however, Prometheus dose not provide such data. One of the solutions to tackle this problem is to cache the first observed value of these metrics as well as the timestamp, then for any subsequent data of the same metric, use the cached timestamp as StartTimestamp and the delta with the first value as value. However, metrics can come and go, or the remote server can restart at any given time, the receiver also needs to take care of issues such as a new value is smaller than the previous seen value, by considering it as a metrics with new StartTime.

Prometheus Metric to OpenTelemetry Metric Proto Mapping

Target as Node

The Target of Prometheus is defined by the scrape_config, it has the information like hostname of the remote service, and a user defined job name which can be used as the service name. These two piece of information makes it a great fit to map it into the Node proto of the OpenTelemetry MetricsData type, as shown below:

type MetricsData struct {
  Node     *commonpb.Node
  Resource *resourcepb.Resource
  Metrics  []*metricspb.Metric
}

The scrape page as whole also can be fit into the above MetricsData data structure, and all the metrics data points can be stored with the Metrics array. We will explain the mappings of individual metirc types in the following couple sections

Metric Value Mapping

In OpenTelemetry, metrics value types can be either int64 or float64, while in Prometheus the value can be safely assumed it's always float64 based on the Prometheus Text Format Document as quoted below:

value is a float represented as required by Go's ParseFloat() function. In addition to standard numerical values, Nan, +Inf, and -Inf are valid values representing not a number, positive infinity, and negative infinity, respectively.

It will make sense for us to stick with this data type as much as possible across all metrics types

Counter

Counter as described in the Prometheus Metric Types Document,

is a cumulative metric that represents a single monotonically increasing counter whose value can only increase or be reset to zero on restart

It is one of the most simple metric types found in both systems, however, it is a cumulative type of metric. Consider what happens when we have two consecutive scrapes from a target, with the first one as shown below:

# HELP http_requests_total The total number of HTTP requests.
# TYPE http_requests_total counter
http_requests_total{method="post",code="200"} 1027
http_requests_total{method="post",code="400"}    3

and the 2nd one:

# HELP http_requests_total The total number of HTTP requests.
# TYPE http_requests_total counter
http_requests_total{method="post",code="200"} 1028
http_requests_total{method="post",code="400"}    5

The Prometheus Receiver will only produce one Metric from the 2nd scrape and subsequent ones if any. The 1st scrape, however, is stored as metadata to calcualate a delta from.

The output of the 2nd scrape is as shown below:

metrics := []*metricspb.Metric{
  {
    MetricDescriptor: &metricspb.MetricDescriptor{
      Name:      "http_requests_total",
      Type:      metricspb.MetricDescriptor_CUMULATIVE_DOUBLE,
      LabelKeys: []*metricspb.LabelKey{{Key: "method"}, {Key: "code"}}},
    Timeseries: []*metricspb.TimeSeries{
      {
        StartTimestamp: startTimestamp,
        LabelValues:    []*metricspb.LabelValue{{Value: "post", HasValue: true}, {Value: "200", HasValue: true}},
        Points: []*metricspb.Point{
          {Timestamp: currentTimestamp, Value: &metricspb.Point_DoubleValue{DoubleValue: 1.0}},
        },
      },
      {
        StartTimestamp: startTimestamp,
        LabelValues:    []*metricspb.LabelValue{{Value: "post", HasValue: false}, {Value: "400", HasValue: true}},
        Points: []*metricspb.Point{
          {Timestamp: currentTimestamp, Value: &metricspb.Point_DoubleValue{DoubleValue: 2.0}},
        },
      },
    },
  },
}

Note: startTimestamp is the timestamp cached from the first scrape, currentTimestamp is the timestamp of the current scrape

Gauge

Gauge, as described in the Prometheus Metric Types Document,

is a metric that represents a single numerical value that can arbitrarily go up and down

# HELP gauge_test some test gauges.
# TYPE gauge_test gague
gauge_test{id="1",foo="bar"} 1.0
gauge_test{id="2",foo=""}    2.0

A major different between Gauges of Prometheus and OpenTelemetry are the value types. In Prometheus, as mentioned earlier, all values can be considered as float type, however, in OpenTelemetry, Gauges can either be Int64 or Double. To make the transformation easier, we always assume the data type is Double.

The corresponding OpenTelemetry Metric of the above examples will be:

metrics := []*metricspb.Metric{
  {
    MetricDescriptor: &metricspb.MetricDescriptor{
      Name:      "gauge_test",
      Type:      metricspb.MetricDescriptor_GAUGE_DOUBLE,
      LabelKeys: []*metricspb.LabelKey{{Key: "id"}, {Key: "foo"}}},
    Timeseries: []*metricspb.TimeSeries{
      {
        StartTimestamp: nil,
        LabelValues:    []*metricspb.LabelValue{{Value: "1", HasValue: true}, {Value: "bar", HasValue: true}},
        Points: []*metricspb.Point{
          {Timestamp: currentTimestamp, Value: &metricspb.Point_DoubleValue{DoubleValue: 1.0}},
        },
      },
      {
        StartTimestamp: nil,
        LabelValues:    []*metricspb.LabelValue{{Value: "2", HasValue: true}, {Value: "", HasValue: false}},
        Points: []*metricspb.Point{
          {Timestamp: currentTimestamp, Value: &metricspb.Point_DoubleValue{DoubleValue: 2.0}},
        },
      },
    },
  },
}

Histogram

Histogram is a complex data type, in Prometheus, it uses multiple data points to represent a single histogram. Its description can be found from: Prometheus Histogram.

Similar to counter, histogram is also a cumulative type metric, thus only the 2nd and subsequent scrapes can produce a metric for OpenTelemetry, with the first scrape stored as metadata.

An example of histogram with first scrape response:

# HELP hist_test This is my histogram vec
# TYPE hist_test histogram
hist_test_bucket{t1="1",,le="10.0"} 1.0
hist_test_bucket{t1="1",le="20.0"} 3.0
hist_test_bucket{t1="1",le="+inf"} 10.0
hist_test_sum{t1="1"} 100.0
hist_test_count{t1="1"} 10.0
hist_test_bucket{t1="2",,le="10.0"} 10.0
hist_test_bucket{t1="2",le="20.0"} 30.0
hist_test_bucket{t1="2",le="+inf"} 100.0
hist_test_sum{t1="2"} 10000.0
hist_test_count{t1="2"} 100.0

And a subsequent 2nd scrape response:

# HELP hist_test This is my histogram vec
# TYPE hist_test histogram
hist_test_bucket{t1="1",,le="10.0"} 2.0
hist_test_bucket{t1="1",le="20.0"} 6.0
hist_test_bucket{t1="1",le="+inf"} 13.0
hist_test_sum{t1="1"} 150.0
hist_test_count{t1="1"} 13.0
hist_test_bucket{t1="2",,le="10.0"} 10.0
hist_test_bucket{t1="2",le="20.0"} 30.0
hist_test_bucket{t1="2",le="+inf"} 100.0
hist_test_sum{t1="2"} 10000.0
hist_test_count{t1="2"} 100.0

Its corresponding OpenTelemetry metrics will be:

metrics := []*metricspb.Metric{
  {
    MetricDescriptor: &metricspb.MetricDescriptor{
      Name:      "hist_test",
      Type:      metricspb.MetricDescriptor_CUMULATIVE_DISTRIBUTION,
      LabelKeys: []*metricspb.LabelKey{{Key: "t1"}}},
    Timeseries: []*metricspb.TimeSeries{
      {
        StartTimestamp: startTimestamp,
        LabelValues:    []*metricspb.LabelValue{{Value: "1", HasValue: true}},
        Points: []*metricspb.Point{
          {Timestamp: currentTimestamp, Value: &metricspb.Point_DistributionValue{
            DistributionValue: &metricspb.DistributionValue{
              BucketOptions: &metricspb.DistributionValue_BucketOptions{
                Type: &metricspb.DistributionValue_BucketOptions_Explicit_{
                  Explicit: &metricspb.DistributionValue_BucketOptions_Explicit{
                    Bounds: []float64{10, 20},
                  },
                },
              },
              Count:   3,
              Sum:     50.0,
              Buckets: []*metricspb.DistributionValue_Bucket{{Count: 1}, {Count: 2}, {Count: 0}},
            }}},
        },
      },
      {
        StartTimestamp: startTimestamp,
        LabelValues:    []*metricspb.LabelValue{{Value: "2", HasValue: true}},
        Points: []*metricspb.Point{
          {Timestamp: currentTimestamp, Value: &metricspb.Point_DistributionValue{
            DistributionValue: &metricspb.DistributionValue{
              BucketOptions: &metricspb.DistributionValue_BucketOptions{
                Type: &metricspb.DistributionValue_BucketOptions_Explicit_{
                  Explicit: &metricspb.DistributionValue_BucketOptions_Explicit{
                    Bounds: []float64{10, 20},
                  },
                },
              },
              Count:   0,
              Sum:     0.0,
              Buckets: []*metricspb.DistributionValue_Bucket{{Count: 0}, {Count: 0}, {Count: 0}},
            }}},
        },
      },
    },
  },
}

There's an important difference between Prometheus bucket and OpenTelemetry bucket that, bucket counts from Prometheus are cumulative, to transform this into OpenTelemetry format, one needs to apply the following formula: CurrentOCBucketVlaue = CurrentPrometheusBucketValue - PrevPrometheusBucketValue

OpenTelemetry does not use +inf as bound, one needs to remove it to generate the Bounds of the OpenTelemetry Bounds.

Other than that, the SumOfSquaredDeviation, which is required by OpenTelemetry format for histogram, is not provided by Prometheus. We have to set this value to 0 instead.

Gaugehistogram

This is an undocumented data type, that's not currently supported.

Summary

Same as histogram, summary is also a complex metric type which is represent by multiple data points. A detailed description can be found from Prometheus Summary

The sum and count from Summary is also cumulative, however, the quantiles are not. The receiver will still consider the first scrape as metadata, and won't produce an output. For any subsequent scrapes, the count and sum will be deltas from the first scrape, while the quantiles are left as it is.

For the following two scrapes, with the first one:

# HELP go_gc_duration_seconds A summary of the GC invocation durations.
# TYPE go_gc_duration_seconds summary
go_gc_duration_seconds{quantile="0"} 0.0001271
go_gc_duration_seconds{quantile="0.25"} 0.0002455
go_gc_duration_seconds{quantile="0.5"} 0.0002904
go_gc_duration_seconds{quantile="0.75"} 0.0003426
go_gc_duration_seconds{quantile="1"} 0.0023638
go_gc_duration_seconds_sum 17.391350544
go_gc_duration_seconds_count 52489

And the 2nd one:

# HELP go_gc_duration_seconds A summary of the GC invocation durations.
# TYPE go_gc_duration_seconds summary
go_gc_duration_seconds{quantile="0"} 0.0001271
go_gc_duration_seconds{quantile="0.25"} 0.0002455
go_gc_duration_seconds{quantile="0.5"} 0.0002904
go_gc_duration_seconds{quantile="0.75"} 0.0003426
go_gc_duration_seconds{quantile="1"} 0.0023639
go_gc_duration_seconds_sum 17.491350544
go_gc_duration_seconds_count 52490

The corresponding OpenTelemetry metrics is as shown below:

metrics := []*metricspb.Metric{
  {
    MetricDescriptor: &metricspb.MetricDescriptor{
      Name:      "go_gc_duration_seconds",
      Type:      metricspb.MetricDescriptor_SUMMARY,
      LabelKeys: []*metricspb.LabelKey{}},
    Timeseries: []*metricspb.TimeSeries{
      {
        StartTimestamp: startTimestamp,
        LabelValues:    []*metricspb.LabelValue{},
        Points: []*metricspb.Point{
          {Timestamp: currentTimestamp, Value: &metricspb.Point_SummaryValue{
            SummaryValue: &metricspb.SummaryValue{
			  Sum:   &wrappers.DoubleValue{Value: 0.1},
			  Count: &wrappers.Int64Value{Value: 1},
              Snapshot: &metricspb.SummaryValue_Snapshot{
                PercentileValues: []*metricspb.SummaryValue_Snapshot_ValueAtPercentile{
                  {Percentile: 0.0,   Value: 0.0001271},
                  {Percentile: 25.0,  Value: 0.0002455},
                  {Percentile: 50.0,  Value: 0.0002904},
                  {Percentile: 75.0,  Value: 0.0003426},
                  {Percentile: 100.0, Value: 0.0023639},
                },
              }}}},
        },
      },
    },
  },
}

There's also some differences between the two systems. One difference is that Prometheus uses quantile, while OpenTelemetry uses percentile. Additionally, OpenTelemetry has optional values for Sum and Count of a snapshot, however, they are not provided by Prometheus, and nil will be used for these values.

Other than that, in some Prometheus client implementations, such as the Python version, Summary is allowed to have no quantiles, in which case the receiver will produce an OpenTelemetry Summary with Snapshot set to nil.

Others

For any other Prometheus metrics types, they will be transformed into the OpenTelemetry Gauge type

Documentation

Overview

Package prometheusreceiver has the logic for scraping Prometheus metrics from already instrumented applications and then passing them onto a metricsink instance.

Index

• func CustomUnmarshalerFunc(v *viper.Viper, viperKey string, sourceViperSection *viper.Viper, ...) error
• type Config
• type Factory
  • func (f *Factory) CreateDefaultConfig() configmodels.Receiver
  • func (f *Factory) CreateMetricsReceiver(logger *zap.Logger, cfg configmodels.Receiver, ...) (component.MetricsReceiver, error)
  • func (f *Factory) CreateTraceReceiver(ctx context.Context, logger *zap.Logger, cfg configmodels.Receiver, ...) (component.TraceReceiver, error)
  • func (f *Factory) CustomUnmarshaler() component.CustomUnmarshaler
  • func (f *Factory) Type() string
• type Preceiver
  • func (pr *Preceiver) Flush()
  • func (pr *Preceiver) Shutdown() error
  • func (pr *Preceiver) Start(host component.Host) error

Functions

func CustomUnmarshalerFunc

func CustomUnmarshalerFunc(v *viper.Viper, viperKey string, sourceViperSection *viper.Viper, intoCfg interface{}) error

CustomUnmarshalerFunc performs custom unmarshaling of config.

Types

type Config

type Config struct {
	configmodels.ReceiverSettings `mapstructure:",squash"`
	PrometheusConfig              *config.Config `mapstructure:"-"`
	BufferPeriod                  time.Duration  `mapstructure:"buffer_period"`
	BufferCount                   int            `mapstructure:"buffer_count"`
	UseStartTimeMetric            bool           `mapstructure:"use_start_time_metric"`

	// ConfigPlaceholder is just an entry to make the configuration pass a check
	// that requires that all keys present in the config actually exist on the
	// structure, ie.: it will error if an unknown key is present.
	ConfigPlaceholder interface{} `mapstructure:"config"`
}

Config defines configuration for Prometheus receiver.

type Factory

type Factory struct {
}

Factory is the factory for receiver.

func (*Factory) CreateDefaultConfig

func (f *Factory) CreateDefaultConfig() configmodels.Receiver

CreateDefaultConfig creates the default configuration for receiver.

func (*Factory) CreateMetricsReceiver

func (f *Factory) CreateMetricsReceiver(
	logger *zap.Logger,
	cfg configmodels.Receiver,
	consumer consumer.MetricsConsumerOld,
) (component.MetricsReceiver, error)

CreateMetricsReceiver creates a metrics receiver based on provided config.

func (*Factory) CreateTraceReceiver

func (f *Factory) CreateTraceReceiver(
	ctx context.Context,
	logger *zap.Logger,
	cfg configmodels.Receiver,
	nextConsumer consumer.TraceConsumerOld,
) (component.TraceReceiver, error)

CreateTraceReceiver creates a trace receiver based on provided config.

func (*Factory) CustomUnmarshaler

func (f *Factory) CustomUnmarshaler() component.CustomUnmarshaler

CustomUnmarshaler returns custom unmarshaler for this config.

func (*Factory) Type

func (f *Factory) Type() string

Type gets the type of the Receiver config created by this factory.

type Preceiver

type Preceiver struct {
	// contains filtered or unexported fields
}

Preceiver is the type that provides Prometheus scraper/receiver functionality.

func (*Preceiver) Flush

func (pr *Preceiver) Flush()

Flush triggers the Flush method on the underlying Prometheus scrapers and instructs them to immediately sned over the metrics they've collected, to the MetricsConsumer. it's not needed on the new prometheus receiver implementation, let it do nothing

func (*Preceiver) Shutdown

func (pr *Preceiver) Shutdown() error

Shutdown stops and cancels the underlying Prometheus scrapers.

func (*Preceiver) Start

func (pr *Preceiver) Start(host component.Host) error

Start is the method that starts Prometheus scraping and it is controlled by having previously defined a Configuration using perhaps New.