Prometheus Flask exporter

PyPI PyPI PyPI - Downloads Travis Coverage Status Code Climate

This library provides HTTP request metrics to export into Prometheus. It can also track method invocations using convenient functions.

Usage

from flask import Flask, request
from prometheus_flask_exporter import PrometheusMetrics

app = Flask(__name__)
metrics = PrometheusMetrics(app)

# static information as metric
metrics.info('app_info', 'Application info', version='1.0.3')

@app.route('/')
def main():
    pass  # requests tracked by default

@app.route('/skip')
@metrics.do_not_track()
def skip():
    pass  # default metrics are not collected

@app.route('/<item_type>')
@metrics.do_not_track()
@metrics.counter('invocation_by_type', 'Number of invocations by type',
         labels={'item_type': lambda: request.view_args['type']})
def by_type(item_type):
    pass  # only the counter is collected, not the default metrics

@app.route('/long-running')
@metrics.gauge('in_progress', 'Long running requests in progress')
def long_running():
    pass

@app.route('/status/<int:status>')
@metrics.do_not_track()
@metrics.summary('requests_by_status', 'Request latencies by status',
                 labels={'status': lambda r: r.status_code})
@metrics.histogram('requests_by_status_and_path', 'Request latencies by status and path',
                   labels={'status': lambda r: r.status_code, 'path': lambda: request.path})
def echo_status(status):
    return 'Status: %s' % status, status

Default metrics

The following metrics are exported by default (unless the export_defaults is set to False).

The prefix for the default metrics can be controlled by the defaults_prefix parameter. If you don't want to use any prefix, pass the prometheus_flask_exporter.NO_PREFIX value in.

To register your own default metrics that will track all registered Flask view functions, use the register_default function.

app = Flask(__name__)
metrics = PrometheusMetrics(app)

@app.route('/simple')
def simple_get():
    pass

metrics.register_default(
    metrics.counter(
        'by_path_counter', 'Request count by request paths',
        labels={'path': lambda: request.path}
    )
)

Note: register your default metrics after all routes have been set up. Also note, that Gauge metrics registered as default will track the /metrics endpoint, and this can't be disabled at the moment.

If you want to apply the same metric to multiple (but not all) endpoints, create its wrapper first, then add to each function.

app = Flask(__name__)
metrics = PrometheusMetrics(app)

by_path_counter = metrics.counter(
    'by_path_counter', 'Request count by request paths',
    labels={'path': lambda: request.path}
)

@app.route('/simple')
@by_path_counter
def simple_get():
    pass

@app.route('/plain')
@by_path_counter
def plain():
    pass

@app.route('/not/tracked/by/path')
def not_tracked_by_path():
    pass

You can avoid recording metrics on individual endpoints by decorating them with @metrics.do_not_track(), or use the excluded_paths argument when creating the PrometheusMetrics instance that takes a regular expression (either a single string, or a list) and matching paths will be excluded. If you have functions that are inherited or otherwise get metrics collected that you don't want, you can use @metrics.exclude_all_metrics() to exclude both default and non-default metrics being collected from it.

Configuration

By default, the metrics are exposed on the same Flask application on the /metrics endpoint and using the core Prometheus registry. If this doesn't suit your needs, set the path argument to None and/or the export_defaults argument to False plus change the registry argument if needed.

The group_by constructor argument controls what the default request duration metric is tracked by: endpoint (function) instead of URI path (the default). This parameter also accepts a function to extract the value from the request, or a name of a property of the request object. Examples:

PrometheusMetrics(app, group_by='path')         # the default
PrometheusMetrics(app, group_by='endpoint')     # by endpoint
PrometheusMetrics(app, group_by='url_rule')     # by URL rule

def custom_rule(req):  # the Flask request object
    """ The name of the function becomes the label name. """
    return '%s::%s' % (req.method, req.path)

PrometheusMetrics(app, group_by=custom_rule)    # by a function

# Error: this is not supported:
PrometheusMetrics(app, group_by=lambda r: r.path)

The group_by_endpoint argument is deprecated since 0.4.0, please use the new group_by argument.

The register_endpoint allows exposing the metrics endpoint on a specific path. It also allows passing in a Flask application to register it on but defaults to the main one if not defined.

Similarly, the start_http_server allows exposing the endpoint on an independent Flask application on a selected HTTP port. It also supports overriding the endpoint's path and the HTTP listen address.

You can also set static labels to add to every request managed by a PrometheusMetrics instance, using the static_labels argument. This needs to be a dictionary, where each key will become a metric label name, and the values their (static) values.

Labels

When defining labels for metrics on functions, the following values are supported in the dictionary:

Label values are evaluated within the request context. The static_labels labels are excepted from this, those need to be static values.

Application information

The PrometheusMetrics.info(..) method provides a way to expose information as a Gauge metric, the application version for example.

The metric is returned from the method to allow changing its value from the default 1:

metrics = PrometheusMetrics(app)
info = metrics.info('dynamic_info', 'Something dynamic')
...
info.set(42.1)

Examples

See some simple examples visualized on a Grafana dashboard by running the demo in the examples/sample-signals folder.

Example dashboard

App Factory Pattern

This library also supports the Flask app factory pattern. Use the init_app method to attach the library to one or more application objects. Note, that to use this mode, you'll need to use the for_app_factory() class method to create the metrics instance, or pass in None for the app in the constructor.

metrics = PrometheusMetrics.for_app_factory()
# then later:
metrics.init_app(app)

Debug mode

Please note, that changes being live-reloaded, when running the Flask app with debug=True, are not going to be reflected in the metrics. See https://github.com/rycus86/prometheus_flask_exporter/issues/4 for more details.

Alternatively - since version 0.5.1 - if you set the DEBUG_METRICS environment variable, you will get metrics for the latest reloaded code. These will be exported on the main Flask app. Serving the metrics on a different port is not going to work most probably - e.g. PrometheusMetrics.start_http_server(..) is not expected to work.

WSGI

Getting accurate metrics for WSGI apps might require a bit more setup. See a working sample app in the examples folder, and also the prometheus_flask_exporter#5 issue.

Multiprocess applications

For multiprocess applications (WSGI or otherwise), you can find some helper classes in the prometheus_flask_exporter.multiprocess module. These provide convenience wrappers for exposing metrics in an environment where multiple copies of the application will run on a single host.

# an extension targeted at Gunicorn deployments
from prometheus_flask_exporter.multiprocess import GunicornPrometheusMetrics

app = Flask(__name__)
metrics = GunicornPrometheusMetrics(app)

# then in the Gunicorn config file:
from prometheus_flask_exporter.multiprocess import GunicornPrometheusMetrics

def when_ready(server):
    GunicornPrometheusMetrics.start_http_server_when_ready(8080)

def child_exit(server, worker):
    GunicornPrometheusMetrics.mark_process_dead_on_child_exit(worker.pid)

Also see the GunicornInternalPrometheusMetrics class if you want to have the metrics HTTP endpoint exposed internally, on the same Flask application.

# an extension targeted at Gunicorn deployments with an internal metrics endpoint
from prometheus_flask_exporter.multiprocess import GunicornInternalPrometheusMetrics

app = Flask(__name__)
metrics = GunicornInternalPrometheusMetrics(app)

# then in the Gunicorn config file:
from prometheus_flask_exporter.multiprocess import GunicornInternalPrometheusMetrics

def child_exit(server, worker):
    GunicornInternalPrometheusMetrics.mark_process_dead_on_child_exit(worker.pid)

There's a small wrapper available for Gunicorn and uWSGI, for everything else you can extend the prometheus_flask_exporter.multiprocess.MultiprocessPrometheusMetrics class and implement the should_start_http_server method at least.

from prometheus_flask_exporter.multiprocess import MultiprocessPrometheusMetrics

class MyMultiprocessMetrics(MultiprocessPrometheusMetrics):
    def should_start_http_server(self):
        return this_worker() == primary_worker()

This should return True on one process only, and the underlying Prometheus client library will collect the metrics for all the forked children or siblings.

Note: this needs the prometheus_multiproc_dir environment variable to point to a valid, writable directory.

You'll also have to call the metrics.start_http_server() function explicitly somewhere, and the should_start_http_server takes care of only starting it once. The examples folder has some working examples on this.

Please also note, that the Prometheus client library does not collect process level metrics, like memory, CPU and Python GC stats when multiprocessing is enabled. See the prometheus_flask_exporter#18 issue for some more context and details.

A final caveat is that the metrics HTTP server will listen on any paths on the given HTTP port, not only on /metrics, and it is not implemented at the moment to be able to change this.

uWSGI lazy-apps

When uWSGI is configured to run with lazy-apps, exposing the metrics endpoint on a separate HTTP server (and port) is not functioning yet. A workaround is to register the endpoint on the main Flask application.

app = Flask(__name__)
metrics = UWsgiPrometheusMetrics(app)
metrics.register_endpoint('/metrics')
# instead of metrics.start_http_server(port)

See #31 for context, and please let me know if you know a better way!

License

MIT