Using the API#


With the Stackdriver Monitoring API, you can work with Stackdriver metric data pertaining to monitored resources in Google Cloud Platform (GCP) or elsewhere.

Essential concepts:

  • Metric data is associated with a monitored resource. A monitored resource has a resource type and a set of resource labels — key-value pairs — that identify the particular resource.
  • A metric further identifies the particular kind of data that is being collected. It has a metric type and a set of metric labels that, when combined with the resource labels, identify a particular time series.
  • A time series is a collection of data points associated with points or intervals in time.

Please refer to the documentation for the Stackdriver Monitoring API for more information.

At present, this client library supports the following features of the API:

  • Querying of time series.
  • Querying of metric descriptors and monitored resource descriptors.
  • Creation and deletion of metric descriptors for custom metrics.
  • Writing of custom metric data.

The Stackdriver Monitoring Client Object#

The Stackdriver Monitoring client library generally makes its functionality available as methods of the monitoring Client class. A Client instance holds authentication credentials and the ID of the target project with which the metric data of interest is associated. This project ID will often refer to a Stackdriver account binding multiple GCP projects and AWS accounts. It can also simply be the ID of a monitored project.

Most often the authentication credentials will be determined implicitly from your environment. See Authentication for more information.

It is thus typical to create a client object as follows:

>>> from import monitoring
>>> client = monitoring.Client(project='target-project')

If you are running in Google Compute Engine or Google App Engine, the current project is the default target project. This default can be further overridden with the GOOGLE_CLOUD_PROJECT environment variable. Using the default target project is even easier:

>>> client = monitoring.Client()

If necessary, you can pass in credentials and project explicitly:

>>> client = monitoring.Client(project='target-project', credentials=...)

Monitored Resource Descriptors#

The available monitored resource types are defined by monitored resource descriptors. You can fetch a list of these with the list_resource_descriptors() method:

>>> for descriptor in client.list_resource_descriptors():
...     print(descriptor.type)

Each ResourceDescriptor has a type, a display name, a description, and a list of LabelDescriptor instances. See the documentation about Monitored Resources for more information.

Metric Descriptors#

The available metric types are defined by metric descriptors. They include platform metrics, agent metrics, and custom metrics. You can list all of these with the list_metric_descriptors() method:

>>> for descriptor in client.list_metric_descriptors():
...     print(descriptor.type)

See MetricDescriptor and the Metric Descriptors API documentation for more information.

You can create new metric descriptors to define custom metrics in the namespace. You do this by creating a MetricDescriptor object using the client’s metric_descriptor() factory and then calling the object’s create() method:

>>> from import MetricKind, ValueType
>>> descriptor = client.metric_descriptor(
...     '',
...     metric_kind=MetricKind.GAUGE,
...     value_type=ValueType.DOUBLE,
...     description='This is a simple example of a custom metric.')
>>> descriptor.create()

You can delete such a metric descriptor as follows:

>>> descriptor = client.metric_descriptor(
...     '')
>>> descriptor.delete()

To define a custom metric parameterized by one or more labels, you must build the appropriate LabelDescriptor objects and include them in the MetricDescriptor object before you call create():

>>> from import LabelDescriptor, LabelValueType
>>> label = LabelDescriptor('response_code', LabelValueType.INT64,
...                         description='HTTP status code')
>>> descriptor = client.metric_descriptor(
...     '',
...     metric_kind=MetricKind.CUMULATIVE,
...     value_type=ValueType.INT64,
...     labels=[label],
...     description='Cumulative count of HTTP responses.')
>>> descriptor.create()


A group is a dynamic collection of monitored resources whose membership is defined by a filter. These groups are usually created via the Stackdriver dashboard. You can list all the groups in a project with the list_groups() method:

>>> for group in client.list_groups():
...     print(, group.display_name, group.parent_id)
('a001', 'Production', None)
('a002', 'Front-end', 'a001')
('1003', 'Back-end', 'a001')

See Group and the API documentation for Groups and Group members for more information.

You can get a specific group based on it’s ID as follows:

>>> group = client.fetch_group('a001')

You can get the current members of this group using the list_members() method:

>>> for member in group.list_members():
...     print(member)

Passing in end_time and start_time to the above method will return historical members based on the current filter of the group. The group membership changes over time, as monitored resources come and go, and as they change properties.

You can create new groups to define new collections of monitored resources. You do this by creating a Group object using the client’s group() factory and then calling the object’s create() method:

>>> filter_string = ' = "us-central1-a"'
>>> group =
...     display_name='My group',
...     filter_string=filter_string,
...     parent_id='a001',
...     is_cluster=True)
>>> group.create()

You can further manipulate an existing group by first initializing a Group object with it’s ID or name, and then calling various methods on it.

Delete a group:

>>> group ='1234')
>>> group.exists()
>>> group.delete()

Update a group:

>>> group ='1234')
>>> group.exists()
>>> group.reload()
>>> group.display_name = 'New Display Name'
>>> group.update()

Time Series Queries#

A time series includes a collection of data points and a set of resource and metric label values. See TimeSeries and the Time Series API documentation for more information.

While you can obtain time series objects by iterating over a Query object, usually it is more useful to retrieve time series data in the form of a pandas.DataFrame, where each column corresponds to a single time series. For this, you must have pandas installed; it is not a required dependency of google-cloud-python.

You can display CPU utilization across your GCE instances over a five minute duration ending at the start of the current minute as follows:

>>> METRIC = ''
>>> query = client.query(METRIC, minutes=5)
>>> print(query.as_dataframe())

Query objects provide a variety of methods for refining the query. You can request temporal alignment and cross-series reduction, and you can filter by label values. See the client query() method and the Query class for more information.

For example, you can display CPU utilization during the last hour across GCE instances with names beginning with "mycluster-", averaged over five-minute intervals and aggregated per zone, as follows:

>>> from import Aligner, Reducer
>>> METRIC = ''
>>> query = (client.query(METRIC, hours=1)
...          .select_metrics(instance_name_prefix='mycluster-')
...          .align(Aligner.ALIGN_MEAN, minutes=5)
...          .reduce(Reducer.REDUCE_MEAN, ''))
>>> print(query.as_dataframe())

Writing Custom Metrics#

The Stackdriver Monitoring API can be used to write data points to custom metrics. Please refer to the documentation on Custom Metrics for more information.

To write a data point to a custom metric, you must provide an instance of Metric specifying the metric type as well as the values for the metric labels. You will need to have either created the metric descriptor earlier (see the Metric Descriptors section) or rely on metric type auto-creation (see Auto-creation of custom metrics).

You will also need to provide a Resource instance specifying a monitored resource type as well as values for all of the monitored resource labels, except for project_id, which is ignored when it’s included in writes to the API. A good choice is to use the underlying physical resource where your application code runs – e.g., a monitored resource type of gce_instance or aws_ec2_instance. In some limited circumstances, such as when only a single process writes to the custom metric, you may choose to use the global monitored resource type.

See Monitored resource types for more information about particular monitored resource types.

>>> from import monitoring
>>> # Create a Resource object for the desired monitored resource type.
>>> resource = client.resource(
...     'gce_instance',
...     labels={
...         'instance_id': '1234567890123456789',
...         'zone': 'us-central1-f'
...     }
... )
>>> # Create a Metric object, specifying the metric type as well as values for any metric labels.
>>> metric = client.metric(
...     type_='',
...     labels={
...         'status': 'successful'
...     }
... )

With a Metric and Resource in hand, the Client can be used to write Point values.

When writing points, the Python type of the value must match the value type of the metric descriptor associated with the metric. For example, a Python float will map to ValueType.DOUBLE.

Stackdriver Monitoring supports several metric kinds: GAUGE, CUMULATIVE, and DELTA. However, DELTA is not supported for custom metrics.

GAUGE metrics represent only a single point in time, so only the end_time should be specified:

>>> client.write_point(metric=metric, resource=resource,
...                    value=3.14, end_time=end_time)  # API call

By default, end_time defaults to utcnow(), so metrics can be written to the current time as follows:

>>> client.write_point(metric, resource, 3.14)  # API call

CUMULATIVE metrics enable the monitoring system to compute rates of increase on metrics that sometimes reset, such as after a process restart. Without cumulative metrics, this reset would otherwise show up as a huge negative spike. For cumulative metrics, the same start time should be re-used repeatedly as more points are written to the time series.

In the examples below, the end_time again defaults to the current time:

>>> RESET = datetime.utcnow()
>>> client.write_point(metric, resource, 3, start_time=RESET)  # API call
>>> client.write_point(metric, resource, 6, start_time=RESET)  # API call

To write multiple TimeSeries in a single batch, you can use write_time_series():

>>> ts1 = client.time_series(metric1, resource, 3.14, end_time=end_time)
>>> ts2 = client.time_series(metric2, resource, 42, end_time=end_time)
>>> client.write_time_series([ts1, ts2])  # API call

While multiple time series can be written in a single batch, each TimeSeries object sent to the API must only include a single point.

All timezone-naive Python datetime objects are assumed to be UTC.