Google Stackdriver Prometheus Exporter

A Prometheus exporter for Google Stackdriver Monitoring metrics. It acts as a proxy that requests Stackdriver API for the metric's time-series everytime prometheus scrapes it.

Installation

Binaries

Download the already existing binaries for your platform:

$ ./stackdriver_exporter <flags>

From source

Using the standard go install (you must have Go already installed in your local machine):

$ go install github.com/prometheus-community/stackdriver_exporter
$ stackdriver_exporter <flags>

Docker

To run the stackdriver exporter as a Docker container, run:

$ docker run -p 9255:9255 prometheuscommunity/stackdriver-exporter <flags>

Kubernetes

You can find a helm chart in the prometheus-community charts repository at https://github.com/prometheus-community/helm-charts/tree/main/charts/prometheus-stackdriver-exporter

helm repo add prometheus-community https://prometheus-community.github.io/helm-charts
helm install [RELEASE_NAME] prometheus-community/prometheus-stackdriver-exporter

Cloud Foundry

The exporter can be deployed to an already existing Cloud Foundry environment:

$ git clone https://github.com/prometheus-community/stackdriver_exporter.git
$ cd stackdriver_exporter

Modify the included application manifest file to include the desired properties. Then you can push the exporter to your Cloud Foundry environment:

$ cf push

BOSH

This exporter can be deployed using the Prometheus BOSH Release.

Usage

Credentials and Permissions

The Google Stackdriver Exporter uses the Google Golang Client Library, which offers a variety of ways to provide credentials. Please refer to the Google Application Default Credentials documentation to see how the credentials can be provided.

If you are using IAM roles, the roles/monitoring.viewer IAM role contains the required permissions. See the Access Control Guide for more information.

If you are still using the legacy Access scopes, the https://www.googleapis.com/auth/monitoring.read scope is required.

Flags

Flag	Required	Default	Description
google.project-id	No	GCloud SDK auto-discovery	Comma seperated list of Google Project IDs
monitoring.metrics-ingest-delay	No		Offsets metric collection by a delay appropriate for each metric type, e.g. because bigquery metrics are slow to appear
monitoring.drop-delegated-projects	No	No	Drop metrics from attached projects and fetch project_id only.
monitoring.metrics-type-prefixes	Yes		Comma separated Google Stackdriver Monitoring Metric Type prefixes (see example and available metrics)
monitoring.metrics-interval	No	5m	Metric's timestamp interval to request from the Google Stackdriver Monitoring Metrics API. Only the most recent data point is used
monitoring.metrics-offset	No	0s	Offset (into the past) for the metric's timestamp interval to request from the Google Stackdriver Monitoring Metrics API, to handle latency in published metrics
monitoring.filters	No		Formatted string to allow filtering on certain metrics type
monitoring.aggregate-deltas	No		If enabled will treat all DELTA metrics as an in-memory counter instead of a gauge. Be sure to read what to know about aggregating DELTA metrics
monitoring.aggregate-deltas-ttl	No	30m	How long should a delta metric continue to be exported and stored after GCP stops producing it. Read slow moving metrics to understand the problem this attempts to solve
monitoring.descriptor-cache-ttl	No	0s	How long should the metric descriptors for a prefixed be cached for
stackdriver.max-retries	No	0	Max number of retries that should be attempted on 503 errors from stackdriver.
stackdriver.http-timeout	No	10s	How long should stackdriver_exporter wait for a result from the Stackdriver API.
stackdriver.max-backoff=	No		Max time between each request in an exp backoff scenario.
stackdriver.backoff-jitter	No	1s	The amount of jitter to introduce in a exp backoff scenario.
stackdriver.retry-statuses	No	503	The HTTP statuses that should trigger a retry.
web.config.file	No		[EXPERIMENTAL] Path to configuration file that can enable TLS or authentication.
web.listen-address	No	:9255	Address to listen on for web interface and telemetry Repeatable for multiple addresses.
web.systemd-socket	No		Use systemd socket activation listeners instead of port listeners (Linux only).
web.stackdriver-telemetry-path	No	"/metrics"	Path under which to expose Stackdriver metrics.
web.telemetry-path	No	/metrics	Path under which to expose Prometheus metrics

TLS and basic authentication

The Stackdriver Exporter supports TLS and basic authentication.

To use TLS and/or basic authentication, you need to pass a configuration file using the --web.config.file parameter. The format of the file is described in the exporter-toolkit repository.

Metrics

The exporter returns the following metrics:

Metric	Description	Labels
stackdriver_monitoring_api_calls_total	Total number of Google Stackdriver Monitoring API calls made	project_id
stackdriver_monitoring_scrapes_total	Total number of Google Stackdriver Monitoring metrics scrapes	project_id
stackdriver_monitoring_scrape_errors_total	Total number of Google Stackdriver Monitoring metrics scrape errors	project_id
stackdriver_monitoring_last_scrape_error	Whether the last metrics scrape from Google Stackdriver Monitoring resulted in an error (1 for error, 0 for success)	project_id
stackdriver_monitoring_last_scrape_timestamp	Number of seconds since 1970 since last metrics scrape from Google Stackdriver Monitoring	project_id
stackdriver_monitoring_last_scrape_duration_seconds	Duration of the last metrics scrape from Google Stackdriver Monitoring	project_id

Metrics gathered from Google Stackdriver Monitoring are converted to Prometheus metrics:

• Metric's names are normalized according to the Prometheus specification using the following pattern:
  1. namespace is a constant prefix (stackdriver)
  2. subsystem is the normalized monitored resource type (ie gce_instance)
  3. name is the normalized metric type (ie compute_googleapis_com_instance_cpu_usage_time)
• Labels attached to each metric are an aggregation of:
  1. the unit in which the metric value is reported
  2. the metric type labels (see Metrics List)
  3. the monitored resource labels (see Monitored Resource Types)
• For each timeseries, only the most recent data point is exported.
• Stackdriver GAUGE metric kinds are reported as Prometheus Gauge metrics
• Stackdriver CUMULATIVE metric kinds are reported as Prometheus Counter metrics.
• Stackdriver DELTA metric kinds are reported as Prometheus Gauge metrics or an accumulating Counter if monitoring.aggregate-deltas is set
• Only BOOL, INT64, DOUBLE and DISTRIBUTION metric types are supported, other types (STRING and MONEY) are discarded.
• DISTRIBUTION metric type is reported as a Prometheus Histogram, except the _sum time series is not supported.

Example

If we want to get all CPU (compute.googleapis.com/instance/cpu) and Disk (compute.googleapis.com/instance/disk) metrics for all Google Compute Engine instances, we can run the exporter with the following options:

stackdriver_exporter \
  --google.project-id=my-test-project \
  --monitoring.metrics-type-prefixes "compute.googleapis.com/instance/cpu,compute.googleapis.com/instance/disk"

Using extra filters:

stackdriver_exporter \
 --google.project-id=my-test-project \
 --monitoring.metrics-type-prefixes='pubsub.googleapis.com/subscription' \
 --monitoring.filters='pubsub.googleapis.com/subscription:resource.labels.subscription_id=monitoring.regex.full_match("us-west4.*my-team-subs.*")'

Filtering enabled collectors

The stackdriver_exporter collects all metrics type prefixes by default.

For advanced uses, the collection can be filtered by using a repeatable URL param called collect. In the Prometheus configuration you can use you can use this syntax under the scrape config.

params:
  collect:
  - compute.googleapis.com/instance/cpu
  - compute.googleapis.com/instance/disk

What to know about Aggregating DELTA Metrics

Treating DELTA Metrics as a gauge produces data which is wildly inaccurate/not very useful (see prometheus-community#116). However, aggregating the DELTA metrics overtime is not a perfect solution and is intended to produce data which mirrors GCP's data as close as possible.

The biggest challenge to producing a correct result is that a counter for prometheus does not start at 0, it starts at the first value which is exported. This can cause inconsistencies when the exporter first starts and for slow moving metrics which are described below.

Start-up Delay

When the exporter first starts it has no persisted counter information and the stores will be empty. When the first sample is received for a series it is intended to be a change from a previous value according to GCP, a delta. But the prometheus counter is not initialized to 0 so it does not export this as a change from 0, it exports that the counter started at the sample value. Since the series exported are dynamic it's not possible to export an initial 0 value in order to account for this issue. The end result is that it can take a few cycles for aggregated metrics to start showing rates exactly as GCP.

As an example consider a prometheus query, sum by(backend_target_name) (rate(stackdriver_https_lb_rule_loadbalancing_googleapis_com_https_request_bytes_count[1m])) which is aggregating 5 series. All 5 series will need to have two samples from GCP in order for the query to produce the same result as GCP.

Slow Moving Metrics

A slow moving metric would be a metric which is not constantly changing with every sample from GCP. GCP does not consistently report slow moving metrics DELTA metrics. If this occurs for too long (default 5m) prometheus will mark the series as stale. The end result is that the next reported sample will be treated as the start of a new series and not an increment from the previous value. Here's an example of this in action,

There are two features which attempt to combat this issue,

1. monitoring.aggregate-deltas-ttl which controls how long a metric is persisted in the data store after its no longer being reported by GCP
2. Metrics which were not collected during a scrape are still exported at their current counter value

The configuration when using monitoring.aggregate-deltas gives a 30 minute buffer to slower moving metrics and monitoring.aggregate-deltas-ttl can be adjusted to tune memory requirements vs correctness. Storing the data for longer results in a higher memory cost.

The feature which continues to export metrics which are not collected can cause the sample has been rejected because another sample with the same timestamp, but a different value, has already been ingested if your scrape config for the exporter has honor_timestamps enabled (this is the default value). This is caused by the fact that it's not possible to know the different between GCP having late arriving data and GCP not exporting a value. The underlying counter is still incremented when this happens so the next reported sample will show a higher rate than expected.

Contributing

Refer to the contributing guidelines.

License

Apache License 2.0, see LICENSE.