API reference

Collectors

An application using the metrics module has 4 primitives, called collectors, at its disposal:

A collector represents one or more observations that change over time.

histogram

A histogram metric is used to collect and analyze statistical data about the distribution of values within the application. Unlike metrics that only allow to track the average value or quantity of events, a histogram allows to see the distribution of values in detail and uncover hidden dependencies.

Suppose there are a lot of individual measurements that you don’t want or can’t store, and the aggregated information (the distribution of values across ranges) is enough to figure out the pattern. In this case, a histogram is used.

Each histogram provides several measurements:

• total count (_count)
• sum of measured values (_sum)
• distribution across buckets (_bucket)

Consider the following problem: you want to know how often the observed value is in the specific range (bucket).

For example, the observed values are 8, 7, 6, 8, 1, 7, 4, 8. Then, in the ranges:

• In the interval [0, 2], there is 1 measurement.
• In the interval [0, 4], there are 2 measurements.
• In the interval [0, 6], there are 3 measurements.
• In the interval [0, infinity], there are 8 measurements (equal to the histogram_demo_count value).

{
    "label_pairs": {
       "le": 2,
       "alias": "my-tnt-app"
    },
    "timestamp": 1680174378390303,
    "metric_name": "histogram_demo_bucket",
    "value": 1
 },
 {
   "label_pairs": {
     "le": 4,
     "alias": "my-tnt-app"
   },
   "timestamp": 1680174378390303,
   "metric_name": "histogram_demo_bucket",
   "value": 2
 },
 {
   "label_pairs": {
     "le": 6,
     "alias": "my-tnt-app"
   },
   "timestamp": 1680174378390303,
   "metric_name": "histogram_demo_bucket",
   "value": 3
 },
 {
   "label_pairs": {
     "le": "inf",
     "alias": "my-tnt-app"
   },
   "timestamp": 1680174378390303,
   "metric_name": "histogram_demo_bucket",
   "value": 8
 },

The metric also displays the count of measurements and their sum:

{
  "label_pairs": {
    "alias": "my-tnt-app"
  },
  "timestamp": 1680180929162484,
  "metric_name": "histogram_demo_count",
  "value": 8
},
{
  "label_pairs": {
    "alias": "my-tnt-app"
  },
  "timestamp": 1680180929162484,
  "metric_name": "histogram_demo_sum",
  "value": 49
},

The design is based on the Prometheus histogram.

metrics.histogram(name[, help, buckets, metainfo])¶

Register a new histogram.

Parameters:	name (string) – collector name. Must be unique. help (string) – collector description. buckets (table) – histogram buckets (an array of sorted positive numbers). The infinity bucket (INF) is appended automatically. Default: {.005, .01, .025, .05, .075, .1, .25, .5, .75, 1.0, 2.5, 5.0, 7.5, 10.0, INF}. metainfo (table) – collector metainfo.
Return:	A histogram object.
Rtype:	histogram_obj

Note

A histogram is basically a set of collectors:

• name .. "_sum" – a counter holding the sum of added observations.
• name .. "_count" – a counter holding the number of added observations.
• name .. "_bucket" – a counter holding all bucket sizes under the label le (less or equal). To access a specific bucket – x (where x is a number), specify the value x for the label le.

object histogram_obj¶

histogram_obj:observe(num, label_pairs)¶

Record a new value in a histogram. This increments all bucket sizes under the labels le >= num and the labels that match label_pairs.

Parameters:	num (number) – value to put in the histogram. label_pairs (table) – table containing label names as keys, label values as values. All internal counters that have these labels specified observe new counter values. Note that both label names and values in label_pairs are treated as strings.

histogram_obj:collect()¶

Return a concatenation of counter_obj:collect() across all internal counters of histogram_obj. For the description of observation, see counter_obj:collect().

histogram_obj:remove(label_pairs)¶

Works like the remove() function of a counter.

summary

A summary metric is used to collect statistical data about the distribution of values within the application.

Each summary provides several measurements:

• total count of measurements
• sum of measured values
• values at specific quantiles

Similar to histograms, a summary also operates with value ranges. However, unlike histograms, it uses quantiles (defined by a number between 0 and 1) for this purpose. In this case, it is not required to define fixed boundaries. For summary type, the ranges depend on the measured values and the number of measurements.

Suppose the example series of measurements are sorted in ascending order: 1, 4, 6, 7, 7, 8, 8, 8.

Quantiles:

• Quantile 0 is the value of the first, minimum element. In this example, it’s 1.
• Quantile 1 is the value of the last, maximum element. In this example, it’s 8.
• Quantile 0.5 is the value of the median element. In this example, it’s 7. This means that the smaller half of the measurements is a range of values from 1 to 7. The larger one is a range of values from 7 to 8.

Note that calculating quantiles requires resources, so it makes sense to calculate no more than one, for example: 0.95 – the majority of measurements.

With a large number of measurements per second, a significant amount of memory is required to store them all. The array is compressed to reduce memory consumption. The degree of compression is determined by an acceptable error rate. In application, error rates mostly from 1% to 10%. This means that a quantile of 0.50 with a 10% error from the example above returns a value in the range of 6.65…7.35 instead of 7.

Additionally, a summary metric doesn’t store values for the whole application’s lifetime. This metric uses a sliding window divided into sections (buckets) where measurements are stored.

Note that buckets in histograms and buckets in quantiles within summaries have different meanings.

local summary_demo = metrics.summary(
    'summary_demo', -- metric name
    'Summary demo', -- description
    {
       [0.5] = 0.01, -- quantile 0.50 with 1% error
       [0.95] = 0.01, -- quantile 0.95 with 1% error
       [0.99] = 0.01, -- quantile 0.99 with 1% error
    },
    {
       max_age_time = 60, -- duration of each bucket in seconds
       age_buckets_count = 5 -- total number of buckets in the sliding window
                             -- window duration = max_age_time * age_buckets_count seconds, or in
                             -- this case = 5 minutes
    }
)

The metric like in the example above returns the following measurements for the specified quantiles:

{
   "label_pairs": {
      "quantile": 0.5,
      "alias": "my-tnt-app"
   },
   "timestamp": 1680180929162484,
   "metric_name": "summary_demo",
   "value": 7
  },
  {
    "label_pairs": {
      "quantile": 0.95,
      "alias": "my-tnt-app"
    },
    "timestamp": 1680180929162484,
    "metric_name": "summary_demo",
    "value": 8
  },
  {
    "label_pairs": {
      "quantile": 0.99,
      "alias": "my-tnt-app"
    },
    "timestamp": 1680180929162484,
    "metric_name": "summary_demo",
    "value": 8
  },

Also, the metric exposes the count of measurements and the sum of observations:

{
  "label_pairs": {
    "alias": "my-tnt-app"
  },
  "timestamp": 1680180929162484,
  "metric_name": "summary_demo_count",
  "value": 8
},
{
  "label_pairs": {
    "alias": "my-tnt-app"
  },
  "timestamp": 1680180929162484,
  "metric_name": "summary_demo_sum",
  "value": 49
},

The design is based on the Prometheus summary.

metrics.summary(name[, help, objectives, params, metainfo])¶

Register a new summary. Quantile computation is based on the “Effective computation of biased quantiles over data streams” algorithm.

Parameters:	name (string) – collector name. Must be unique. help (string) – collector description. objectives (table) – a list of “targeted” φ-quantiles in the {quantile = error, ... } form. Example: {[0.5]=0.01, [0.9]=0.01, [0.99]=0.01}. The targeted φ-quantile is specified in the form of a φ-quantile and the tolerated error. For example, {[0.5] = 0.1} means that the median (= 50th percentile) is to be returned with a 10-percent error. Note that percentiles and quantiles are the same concept, except that percentiles are expressed as percentages. The φ-quantile must be in the interval [0, 1]. A lower tolerated error for a φ-quantile results in higher memory and CPU usage during summary calculation. params (table) – table of the summary parameters used to configuring the sliding time window. This window consists of several buckets to store observations. New observations are added to each bucket. After a time period, the head bucket (from which observations are collected) is reset, and the next bucket becomes the new head. This way, each bucket stores observations for max_age_time * age_buckets_count seconds before it is reset. max_age_time sets the duration of each bucket’s lifetime – that is, how many seconds the observations are kept before they are discarded. age_buckets_count sets the number of buckets in the sliding time window. This variable determines the number of buckets used to exclude observations older than max_age_time from the summary. The value is a trade-off between resources (memory and CPU for maintaining the bucket) and how smooth the time window moves. Default value: {max_age_time = math.huge, age_buckets_count = 1}. metainfo (table) – collector metainfo.
Return:	A summary object.
Rtype:	summary_obj

Note

A summary represents a set of collectors:

• name .. "_sum" – a counter holding the sum of added observations.
• name .. "_count" – a counter holding the number of added observations.
• name holds all the quantiles under observation that find themselves under the label quantile (less or equal). To access bucket x (where x is a number), specify the value x for the label quantile.

object summary_obj¶

summary_obj:observe(num, label_pairs)¶

Record a new value in a summary.

Parameters:

num (number) – value to put in the data stream. label_pairs (table) – a table containing label names as keys, label values as values. All internal counters that have these labels specified observe new counter values. You can’t add the "quantile" label to a summary. It is added automatically. If max_age_time and age_buckets_count are set, the observed value is added to each bucket. Note that both label names and values in label_pairs are treated as strings.

summary_obj:collect()¶

Return a concatenation of counter_obj:collect() across all internal counters of summary_obj. For the description of observation, see counter_obj:collect(). If max_age_time and age_buckets_count are set, quantile observations are collected only from the head bucket in the sliding time window, not from every bucket. If no observations were recorded, the method will return NaN in the values.

summary_obj:remove(label_pairs)¶

Works like the remove() function of a counter.

Labels

All collectors support providing label_pairs on data modification. A label is a piece of metainfo that you associate with a metric in the key-value format. For details, see tags in Graphite and labels in Prometheus.

Labels are used to differentiate between the characteristics of a thing being measured. For example, in a metric associated with the total number of HTTP requests, you can represent methods and statuses as label pairs:

http_requests_total_counter:inc(1, {method = 'POST', status = '200'})

You don’t have to predefine labels in advance.

With labels, you can extract new time series (visualize their graphs) by specifying conditions with regard to label values. The example above allows extracting the following time series:

1. The total number of requests over time with method = "POST" (and any status).
2. The total number of requests over time with status = 500 (and any method).

You can also set global labels by calling metrics.set_global_labels({ label = value, ...}).

Metrics functions

metrics.cfg([config])¶

Entrypoint to setup the module. Since 0.17.0.

Parameters:

config (table) – module configuration options: cfg.include (string/table, default 'all'): 'all to enable all supported default metrics, 'none' to disable all default metrics, table with names of the default metrics to enable a specific set of metrics. cfg.exclude (table, default {}): table containing the names of the default metrics that you want to disable. Has higher priority than cfg.include. cfg.labels (table, default {}): table containing label names as string keys, label values as values.

You can work with metrics.cfg as a table to read values, but you must call metrics.cfg{} as a function to update them.

Supported default metric names (for cfg.include and cfg.exclude tables):

• all (metasection including all metrics)
• network
• operations
• system
• replicas
• info
• slab
• runtime
• memory
• spaces
• fibers
• cpu
• vinyl
• memtx
• luajit
• cartridge_issues
• cartridge_failover
• clock
• event_loop

See metrics reference for details. All metric collectors from the collection have metainfo.default = true.

cfg.labels are the global labels to be added to every observation.

Global labels are applied only to metric collection. They have no effect on how observations are stored.

Global labels can be changed on the fly.

label_pairs from observation objects have priority over global labels. If you pass label_pairs to an observation method with the same key as some global label, the method argument value will be used.

Note that both label names and values in label_pairs are treated as strings.

metrics.enable_default_metrics([include, exclude])¶

Same as metrics.cfg{include=include, exclude=exclude}, but include={} is treated as include='all' for backward compatibility.

metrics.set_global_labels(label_pairs)¶

Same as metrics.cfg{labels=label_pairs}.

metrics.collect([opts])¶

Collect observations from each collector.

Parameters:	opts (table) – table of collect options: invoke_callbacks – if true, invoke_callbacks() is triggered before actual collect. default_only – if true, observations contain only default metrics (metainfo.default = true).

object registry¶

registry:unregister(collector)¶

Remove a collector from the registry.

Parameters:	collector (collector_obj) – the collector to be removed.

Example:

local collector = metrics.gauge('some-gauge')

-- after a while, we don't need it anymore

metrics.registry:unregister(collector)

registry:find(kind, name)¶

Find a collector in the registry.

Parameters:	kind (string) – collector kind (counter, gauge, histogram, or summary). name (string) – collector name.
Return:	A collector object or nil.
Rtype:	collector_obj

Example:

local collector = metrics.gauge('some-gauge')

collector = metrics.registry:find('gauge', 'some-gauge')

metrics.register_callback(callback)¶

Register a function named callback, which will be called right before metric collection on plugin export.

Parameters:	callback (function) – a function that takes no parameters.

This method is most often used for gauge metrics updates.

Example:

metrics.register_callback(function()
    local cpu_metrics = require('metrics.psutils.cpu')
    cpu_metrics.update()
end)

metrics.unregister_callback(callback)¶

Unregister a function named callback that is called right before metric collection on plugin export.

Parameters:	callback (function) – a function that takes no parameters.

Example:

local cpu_callback = function()
    local cpu_metrics = require('metrics.psutils.cpu')
    cpu_metrics.update()
end

metrics.register_callback(cpu_callback)

-- after a while, we don't need that callback function anymore

metrics.unregister_callback(cpu_callback)

metrics.invoke_callbacks()¶

Invoke all registered callbacks. Has to be called before each collect(). (Since version 0.16.0, you may use collect{invoke_callbacks = true} instead.) If you’re using one of the default exporters, invoke_callbacks() will be called by the exporter.

Metrics role API

Below are the functions that you can call with metrics = require('cartridge.roles.metrics') specified in your init.lua.

metrics.set_export(export)¶

Parameters:	export (table) – a table containing paths and formats of the exported metrics.

Configure the endpoints of the metrics role:

local metrics = require('cartridge.roles.metrics')
metrics.set_export({
    {
        path = '/path_for_json_metrics',
        format = 'json'
    },
    {
        path = '/path_for_prometheus_metrics',
        format = 'prometheus'
    },
    {
        path = '/health',
        format = 'health'
    }
})

You can add several entry points of the same format but with different paths, for example:

metrics.set_export({
    {
        path = '/path_for_json_metrics',
        format = 'json'
    },
    {
        path = '/another_path_for_json_metrics',
        format = 'json'
    },
})

metrics.set_default_labels(label_pairs)¶

Add default global labels. Note that both label names and values in label_pairs are treated as strings.

Parameters:	label_pairs (table) – Table containing label names as string keys, label values as values.

local metrics = require('cartridge.roles.metrics')
metrics.set_default_labels({ ['my-custom-label'] = 'label-value' })

Collecting HTTP request latency statistics

The metrics module provides middleware for monitoring HTTP (set by the http module) latency statistics. The latency collector observes both latency information and the number of invocations. The metrics collected by HTTP middleware are separated by a set of labels:

• route (path)
• method (method)
• HTTP status code (status)

For each route that you want to track, you must specify the middleware explicitly. The middleware does not cover the 404 errors.

metrics.http_middleware.configure_default_collector(type_name, name, help)¶

Register a collector for the middleware and set it as default.

Parameters:	type_name (string) – collector type: histogram or summary. The default is histogram. name (string) – collector name. The default is http_server_request_latency. help (string) – collector description. The default is HTTP Server Request Latency.

Possible errors:

• A collector with the same type and name already exists in the registry.

metrics.http_middleware.build_default_collector(type_name, name[, help])¶

Register and return a collector for the middleware.

Parameters:	type_name (string) – collector type: histogram or summary. The default is histogram. name (string) – collector name. The default is http_server_request_latency. help (string) – collector description. The default is HTTP Server Request Latency.
Return:	A collector object.

Possible errors:

• A collector with the same type and name already exists in the registry.

metrics.http_middleware.set_default_collector(collector)¶

Set the default collector.

Parameters:	collector – middleware collector object.

metrics.http_middleware.get_default_collector()¶

Return the default collector. If the default collector hasn’t been set yet, register it (with default http_middleware.build_default_collector(...) parameters) and set it as default.

Return:	A collector object.

metrics.http_middleware.v1(handler, collector)¶

Latency measuring wrap-up for the HTTP ver. 1.x.x handler. Returns a wrapped handler.

Parameters:	handler (function) – handler function. collector – middleware collector object. If not set, the default collector is used (like in http_middleware.get_default_collector()).

Usage: httpd:route(route, http_middleware.v1(request_handler, collector))

Example:

#!/usr/bin/env tarantool
package.path = package.path .. ";../?.lua"

local json = require('json')
local fiber = require('fiber')
local metrics = require('metrics')
local log = require('log')
local http_middleware = metrics.http_middleware

-- Configure HTTP routing
local ip = '127.0.0.1'
local port = 12345
local httpd = require('http.server').new(ip, port) -- HTTP ver. 1.x.x
local route = { path = '/path', method = 'POST' }

-- Route handler
local handler = function(req)
    for _ = 1, 10 do
        fiber.sleep(0.1)
    end

    return { status = 200, body = req.body }
end

-- Configure summary latency collector
local collector = http_middleware.build_default_collector('summary')

-- Set route handler with summary latency collection
httpd:route(route, http_middleware.v1(handler, collector))
-- Start HTTP routing
httpd:start()

-- Set HTTP client, make some request
local http_client = require("http.client") -- HTTP ver. 1.x.x
http_client.post('http://' .. ip .. ':' .. port .. route.path, json.encode({ body = 'text' }))

-- Collect the metrics
log.info(metrics.collect())
--[[

- label_pairs:
    path: /path
    method: POST
    status: 200
  timestamp: 1588951616500768
  value: 1
  metric_name: path_latency_count

- label_pairs:
    path: /path
    method: POST
    status: 200
  timestamp: 1588951616500768
  value: 1.0240110000595
   metric_name: path_latency_sum

 - label_pairs:
     path: /path
     method: POST
     status: 200
     quantile: 0.5
   timestamp: 1588951616500768
   value: 1.0240110000595
   metric_name: path_latency

 - label_pairs:
     path: /path
     method: POST
     status: 200
     quantile: 0.9
   timestamp: 1588951616500768
   value: 1.0240110000595
   metric_name: path_latency

 - label_pairs:
     path: /path
     method: POST
     status: 200
     quantile: 0.99
   timestamp: 1588951616500768
   value: 1.0240110000595
   metric_name: path_latency

--]]

-- Exit event loop
os.exit()

CPU usage metrics

CPU metrics work only on Linux. See the metrics reference for details.

To enable CPU metrics, first register a callback function:

local metrics = require('metrics')

local cpu_callback = function()
    local cpu_metrics = require('metrics.psutils.cpu')
    cpu_metrics.update()
end

metrics.register_callback(cpu_callback)

Collected metrics example:

# HELP tnt_cpu_time Host CPU time
# TYPE tnt_cpu_time gauge
tnt_cpu_time 15006759
# HELP tnt_cpu_thread Tarantool thread cpu time
# TYPE tnt_cpu_thread gauge
tnt_cpu_thread{thread_name="coio",file_name="init.lua",thread_pid="699",kind="system"} 160
tnt_cpu_thread{thread_name="tarantool",file_name="init.lua",thread_pid="1",kind="user"} 949
tnt_cpu_thread{thread_name="tarantool",file_name="init.lua",thread_pid="1",kind="system"} 920
tnt_cpu_thread{thread_name="coio",file_name="init.lua",thread_pid="11",kind="user"} 79
tnt_cpu_thread{thread_name="coio",file_name="init.lua",thread_pid="699",kind="user"} 44
tnt_cpu_thread{thread_name="coio",file_name="init.lua",thread_pid="11",kind="system"} 294

Prometheus query aggregated by thread name:

sum by (thread_name) (idelta(tnt_cpu_thread[$__interval]))
  / scalar(idelta(tnt_cpu_total[$__interval]) / tnt_cpu_count)

All psutils metric collectors have metainfo.default = true.

To clear CPU metrics when you don’t need them anymore, remove the callback and clear the collectors with a method:

metrics.unregister_callback(cpu_callback)
cpu_metrics.clear()

Examples

Below are some examples of using metric primitives.

Notice that this usage is independent of export plugins such as Prometheus, Graphite, etc. For documentation on how to use the plugins, see the Metrics plugins section.

Using counters:

local metrics = require('metrics')

-- create a counter
local http_requests_total_counter = metrics.counter('http_requests_total')

-- somewhere in the HTTP requests middleware:
http_requests_total_counter:inc(1, {method = 'GET'})

Using gauges:

local metrics = require('metrics')

-- create a gauge
local cpu_usage_gauge = metrics.gauge('cpu_usage', 'CPU usage')

-- register a lazy gauge value update
-- this will be called whenever export is invoked in any plugins
metrics.register_callback(function()
    local current_cpu_usage = some_cpu_collect_function()
    cpu_usage_gauge:set(current_cpu_usage, {app = 'tarantool'})
end)

Using histograms:

local metrics = require('metrics')
local fiber = require('fiber')
-- create a histogram
local http_requests_latency_hist = metrics.histogram(
    'http_requests_latency', 'HTTP requests total', {2, 4, 6})

-- somewhere in the HTTP request middleware:

local t0 = fiber.clock()
observable_function()
local t1 = fiber.clock()

local latency = t1 - t0
http_requests_latency_hist:observe(latency)

Using summaries:

local metrics = require('metrics')
local fiber = require('fiber')

-- create a summary with a window of 5 age buckets and a bucket lifetime of 60 s
local http_requests_latency = metrics.summary(
    'http_requests_latency', 'HTTP requests total',
    {[0.5]=0.01, [0.9]=0.01, [0.99]=0.01},
    {max_age_time = 60, age_buckets_count = 5}
)

-- somewhere in the HTTP requests middleware:
local t0 = fiber.clock()
observable_function()
local t1 = fiber.clock()

local latency = t1 - t0
http_requests_latency:observe(latency)