Amazon CloudWatch Construct Library

--- ![Stability: Stable]( ---

Metric objects

Metric objects represent a metric that is emitted by AWS services or your own application, such as CPUUsage, FailureCount or Bandwidth.

Metric objects can be constructed directly or are exposed by resources as attributes. Resources that expose metrics will have functions that look like metricXxx() which will return a Metric object, initialized with defaults that make sense.

For example, lambda.Function objects have the fn.metricErrors() method, which represents the amount of errors reported by that Lambda function:

# Example automatically generated without compilation. See
errors = fn.metric_errors()

Instantiating a new Metric object

If you want to reference a metric that is not yet exposed by an existing construct, you can instantiate a Metric object to represent it. For example:

# Example automatically generated without compilation. See
metric = Metric(
        "ProcessingStep": "Download"

Metric Math

Math expressions are supported by instantiating the MathExpression class. For example, a math expression that sums two other metrics looks like this:

# Example automatically generated without compilation. See
all_problems = MathExpression(
    expression="errors + faults",
        "errors": my_construct.metric_errors(),
        "faults": my_construct.metric_faults()

You can use MathExpression objects like any other metric, including using them in other math expressions:

# Example automatically generated without compilation. See
problem_percentage = MathExpression(
    expression="(problems / invocations) * 100",
        "problems": all_problems,
        "invocations": my_construct.metric_invocations()


To graph or alarm on metrics you must aggregate them first, using a function like Average or a percentile function like P99. By default, most Metric objects returned by CDK libraries will be configured as Average over 300 seconds (5 minutes). The exception is if the metric represents a count of discrete events, such as failures. In that case, the Metric object will be configured as Sum over 300 seconds, i.e. it represents the number of times that event occurred over the time period.

If you want to change the default aggregation of the Metric object (for example, the function or the period), you can do so by passing additional parameters to the metric function call:

# Example automatically generated without compilation. See
minute_error_rate = fn.metric_errors(
    label="Lambda failure rate"

This function also allows changing the metric label or color (which will be useful when embedding them in graphs, see below).

Rates versus Sums

The reason for using Sum to count discrete events is that some events are emitted as either 0 or 1 (for example Errors for a Lambda) and some are only emitted as 1 (for example NumberOfMessagesPublished for an SNS topic).

In case 0-metrics are emitted, it makes sense to take the Average of this metric: the result will be the fraction of errors over all executions.

If 0-metrics are not emitted, the Average will always be equal to 1, and not be very useful.

In order to simplify the mental model of Metric objects, we default to aggregating using Sum, which will be the same for both metrics types. If you happen to know the Metric you want to alarm on makes sense as a rate (Average) you can always choose to change the statistic.


Alarms can be created on metrics in one of two ways. Either create an Alarm object, passing the Metric object to set the alarm on:

# Example automatically generated without compilation. See
Alarm(self, "Alarm",

Alternatively, you can call metric.createAlarm():

# Example automatically generated without compilation. See
fn.metric_errors().create_alarm(self, "Alarm",

The most important properties to set while creating an Alarms are:

  • threshold: the value to compare the metric against.

  • comparisonOperator: the comparison operation to use, defaults to metric >= threshold.

  • evaluationPeriods: how many consecutive periods the metric has to be breaching the the threshold for the alarm to trigger.

A note on units

In CloudWatch, Metrics datums are emitted with units, such as seconds or bytes. When Metric objects are given a unit attribute, it will be used to filter the stream of metric datums for datums emitted using the same unit attribute.

In particular, the unit field is not used to rescale datums or alarm threshold values (for example, it cannot be used to specify an alarm threshold in Megabytes if the metric stream is being emitted as bytes).

You almost certainly don’t want to specify the unit property when creating Metric objects (which will retrieve all datums regardless of their unit), unless you have very specific requirements. Note that in any case, CloudWatch only supports filtering by unit for Alarms, not in Dashboard graphs.

Please see the following GitHub issue for a discussion on real unit calculations in CDK:


Dashboards are set of Widgets stored server-side which can be accessed quickly from the AWS console. Available widgets are graphs of a metric over time, the current value of a metric, or a static piece of Markdown which explains what the graphs mean.

The following widgets are available:

  • GraphWidget – shows any number of metrics on both the left and right vertical axes.

  • AlarmWidget – shows the graph and alarm line for a single alarm.

  • SingleValueWidget – shows the current value of a set of metrics.

  • TextWidget – shows some static Markdown.

Graph widget

A graph widget can display any number of metrics on either the left or right vertical axis:

# Example automatically generated without compilation. See
    title="Executions vs error rate",


        label="Error rate",

Alarm widget

An alarm widget shows the graph and the alarm line of a single alarm:

# Example automatically generated without compilation. See

Single value widget

A single-value widget shows the latest value of a set of metrics (as opposed to a graph of the value over time):

# Example automatically generated without compilation. See
    metrics=[visitor_count, purchase_count]

Text widget

A text widget shows an arbitrary piece of MarkDown. Use this to add explanations to your dashboard:

# Example automatically generated without compilation. See
    markdown="# Key Performance Indicators"

Dashboard Layout

The widgets on a dashboard are visually laid out in a grid that is 24 columns wide. Normally you specify X and Y coordinates for the widgets on a Dashboard, but because this is inconvenient to do manually, the library contains a simple layout system to help you lay out your dashboards the way you want them to.

Widgets have a width and height property, and they will be automatically laid out either horizontally or vertically stacked to fill out the available space.

Widgets are added to a Dashboard by calling add(widget1, widget2, ...). Widgets given in the same call will be laid out horizontally. Widgets given in different calls will be laid out vertically. To make more complex layouts, you can use the following widgets to pack widgets together in different ways:

  • Column: stack two or more widgets vertically.

  • Row: lay out two or more widgets horizontally.

  • Spacer: take up empty space