Class: Frankenstein::Request

Inherits:
Object
  • Object
show all
Defined in:
lib/frankenstein/request.rb

Overview

A common pattern for statistical instrumentation is to capture a few basic numbers for all incoming and outgoing requests to the service. Since this is a common pattern, we can abstract that behaviour into a common class, which simplifies the external interface for maintaining statistics in this common case.

For more information on this pattern, see https://honeycomb.io/blog/2017/01/instrumentation-the-first-four-things-you-measure/

Defined Under Namespace

Classes: NoBlockError

Instance Method Summary collapse

Constructor Details

#initialize(prefix, labels: [], duration_labels: nil, outgoing: true, description: prefix, registry: Prometheus::Client.registry) ⇒ Request

Create a new request instrumentation package.

A "request", for the purposes of this discussion, is a distinct interaction with an external system, typically either the receipt of some sort of communication from another system which needs a response by this system (the one being instrumented), or else the communication to another system from this one for which we are expecting an answer. Each instance of this class should be used to instrument all requests of a particular type.

For each instance of this class, the following metrics will be created:

  • <prefix>_requests_total -- a counter indicating the total number of requests started (initiated or received by the system). Labels on this metric are taken from the label set passed to #measure.

  • <prefix>_request_duration_seconds -- a histogram for the response times of successful responses (that is, where no exception was raised). You can get the count of total successful responses from <prefix>_request_duration_seconds_count. Labels on this metric are taken from the labels set generated during the measured run (as generated by manipulating the hash yielded to your block).

  • <prefix>_exceptions_total -- a count of the number of exceptions raised during processing. A label, class, indicates the class of the exception raised. Labels on this metric are taken from the label set passed to #measure, along with a special label class to indicate the class of the exception raised.

  • <prefix>_in_progress_count -- a gauge indicating how many requests are currently in progress as at the time of the scrape. Labels on this metric are taken from the label set passed to #measure.

Parameters:

  • prefix (#to_s) —

    the string that will be prepended to all of the Prometheus metric names generated for this instrumentation. The prefix you choose should include both the application name (typically the first word) as well as a unique identifier for the request type itself. Multiple words should be underscore separated.

  • outgoing (Boolean) (defaults to: true) —

    whether this Request instance is collecting data on incoming requests or outgoing requests (the default, as usually there is one incoming request handler, but there can easily be several outgoing request types). It is only used to customise the metric description text for the metrics, so it's not crucially important.

  • description (#to_s) (defaults to: prefix) —

    a short explanation of what this is measuring. It should be a singular and indefinite noun phrase, to maximise the chances that it will fit neatly into the generated description text.

  • labels (Array<Symbol>) (defaults to: []) —

    the names of the labels that will be set on most of the metrics created by this instance.

  • duration_labels (Array<Symbol>) (defaults to: nil) —

    the names of the labels which will be set on the _duration_seconds histogram. By default, the value of the labels parameter will be used, but if you want to have a separate label set for your duration histogram, this is what you want.

  • registry (Prometheus::Client::Registry) (defaults to: Prometheus::Client.registry) —

    the client registry in which all the metrics will be created. The default will put all the metrics in the Prometheus Client's default registry, which may or may not be what you're up for. If you're using Frankenstein::Server, you want stats_server.registry.



82
83
84
85
86
87
88
89
90
91
92
93
94
 
# File 'lib/frankenstein/request.rb', line 82

def initialize(prefix, labels: [], duration_labels: nil, outgoing: true, description: prefix, registry: Prometheus::Client.registry)
  @requests   = registry.counter(:"#{prefix}_requests_total", docstring: "Number of #{description} requests #{outgoing ? 'sent' : 'received'}", labels: labels)
  @durations  = registry.histogram(:"#{prefix}_request_duration_seconds", docstring: "Time taken to #{outgoing ? 'receive' : 'send'} a #{description} response", labels: duration_labels || labels)
  @exceptions = registry.counter(:"#{prefix}_exceptions_total", docstring: "Number of exceptions raised by the #{description} code", labels: labels + [:class])
  @current    = registry.gauge(:"#{prefix}_in_progress_count", docstring: "Number of #{description} requests currently in progress", labels: labels)

  # Prometheus::Client::Gauge doesn't (yet) have a built-in way to
  # atomically "adjust" a gauge, only get the current value and set a
  # new value.  To avoid the resulting textbook race condition, we
  # need to wrap the get/set pair of operations in this handy-dandy
  # mutex.
  @mutex = Mutex.new
end

Instance Method Details

#measure(labels = {}) {|Hash| ... } ⇒ Object

Instrument an instance of the request.

Each time a particular external communication occurs, it should be wrapped by a call to this method. Request-related statistics (that the request has been made or received) are updated before the passed block is executed, and then after the block completes, response-related statistics (duration or exception) are recorded. The number of currently-in-progress instances of the request are also kept track of.

Parameters:

  • labels (Hash) (defaults to: {}) —

    a set of labels that can help to differentiate different sorts of requests. These labels are applied to the <prefix>_requests_total and <prefix>_in_progress_count metrics, as well as the <prefix>_exceptions_total metric, if an exception is raised.

    Don't get too fancy with this label set -- it's unusual that this is actually useful in practice. However it is provided for those unusual cases where it isn't a bad idea. Your go-to solution should be to label the <prefix>_request_duration_seconds metric (by modifying the hash yielded to the block you pass to #measure), rather than using this parameter with wild abandon.

    Serious talk time: I've been there. It seems like a great idea at first, to differentiate requests with lots of labels, but it usually just ends up turning into a giant mess. Primarily, due to the way that Prometheus deals with label sets, if you ever use a label on any of your requests, you need to set the same label to some value on all of your requests. So, unless you can say with certainty that every request you receive will logically have some meaningful value for a given label, you shouldn't use it.

    NOTE: the labelset you specify here will be the default labelset applied to the <prefix>_request_duration_seconds metric. If you need to remove a label from the response, use labels.replace or labels.delete to remove the key.

Yields:

  • (Hash) —

    the labels that will be applied to the <Prefix>_request_duration_seconds metric.

    In order for your label set to be applied, you must mutate the hash that is yielded, rather than overwriting it. That means, for example, that the following code will not work:

    req_stats.measure do |labels|
  labels = {foo: 'bar', baz: 'wombat'}
  ...

    Instead, you need to either set each key one by one, or use the handy-dandy Hash#replace method, like this:

    req_stats.measure do |labels|
  labels.replace(foo: 'bar', baz: 'wombat')
  ...

    If your labels are not being applied to your response histogram, check for any assignment to the yielded variable. It's really easy to do by mistake.

    NOTE WELL: The Prometheus specification (assuming it exists) apparently requires that all of the instances of a given metric have the same set of labels. If you fail to do this, an exception will be raised by Prometheus after the block is executed.

Returns:

  • (Object) —

    whatever was returned by the block passed.

Raises:

  • (Frankenstein::Request::NoBlockError) —

    if you didn't pass a block to call. There's nothing to instrument!

  • (Prometheus::Client::LabelSetValidator::LabelSetError) —

    if you violate any written or unwritten rules about how Prometheus label sets should be constructed.

  • (Exception) —

    any exception raised by the executed block will be re-raised by this method after statistics collection is complete.



173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
 
# File 'lib/frankenstein/request.rb', line 173

def measure(labels = {})
  start_time = Time.now

  unless block_given?
    raise NoBlockError,
      "No block passed to #{self.class}#measure"
  end

  @requests.increment(labels: labels)
  @mutex.synchronize { @current.set((@current.get(labels: labels) || 0) + 1, labels: labels) }

  res_labels = labels.dup

  begin
    yield(res_labels).tap do
      elapsed_time = Time.now - start_time
      @durations.observe(elapsed_time, labels: res_labels)
    end
  rescue Exception => ex
    @exceptions.increment(labels: labels.merge(class: ex.class.to_s))
    raise
  ensure
    @mutex.synchronize { @current.set(@current.get(labels: labels) - 1, labels: labels) }
  end
end