Important: This documentation is about an older version. It's relevant only to the release noted, many of the features and functions have been updated or replaced. Please view the current version.

Documentation Grafana Agent Flow mode Reference Components reference prometheus.remote_write

prometheus.remote_write

prometheus.remote_write collects metrics sent from other components into a Write-Ahead Log (WAL) and forwards them over the network to a series of user-supplied endpoints. Metrics are sent over the network using the Prometheus Remote Write protocol.

Multiple prometheus.remote_write components can be specified by giving them different labels.

Usage

river Copy

prometheus.remote_write "LABEL" {
  endpoint {
    url = REMOTE_WRITE_URL

    ...
  }

  ...
}

Arguments

The following arguments are supported:

Blocks

The following blocks are supported inside the definition of prometheus.remote_write:

The > symbol indicates deeper levels of nesting. For example, endpoint > basic_auth refers to a basic_auth block defined inside an endpoint block.

endpoint block

The endpoint block describes a single location to send metrics to. Multiple endpoint blocks can be provided to send metrics to multiple locations.

The following arguments are supported:

At most one of the following can be provided:

• bearer_token argument.
• bearer_token_file argument.
• basic_auth block.
• authorization block.
• oauth2 block.
• sigv4 block.
• azuread block.

When multiple endpoint blocks are provided, metrics are concurrently sent to all configured locations. Each endpoint has a queue which is used to read metrics from the WAL and queue them for sending. The queue_config block can be used to customize the behavior of the queue.

Endpoints can be named for easier identification in debug metrics using the name argument. If the name argument isn’t provided, a name is generated based on a hash of the endpoint settings.

When send_native_histograms is true, native Prometheus histogram samples sent to prometheus.remote_write are forwarded to the configured endpoint. If the endpoint doesn’t support receiving native histogram samples, pushing metrics fails.

basic_auth block

password and password_file are mutually exclusive, and only one can be provided inside a basic_auth block.

authorization block

credential and credentials_file are mutually exclusive, and only one can be provided inside an authorization block.

oauth2 block

client_secret and client_secret_file are mutually exclusive, and only one can be provided inside an oauth2 block.

The oauth2 block may also contain a separate tls_config sub-block.

sigv4 block

If region is left blank, the region from the default credentials chain is used.

If access_key is left blank, the environment variable AWS_ACCESS_KEY_ID is used.

If secret_key is left blank, the environment variable AWS_SECRET_ACCESS_KEY is used.

azuread block

The supported values for cloud are:

• "AzurePublic"
• "AzureChina"
• "AzureGovernment"

managed_identity block

client_id should be a valid UUID in one of the supported formats:

• xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx
• urn:uuid:xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx
• Microsoft encoding: {xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx}
• Raw hex encoding: xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

tls_config block

The following pairs of arguments are mutually exclusive and can’t both be set simultaneously:

• ca_pem and ca_file
• cert_pem and cert_file
• key_pem and key_file

When configuring client authentication, both the client certificate (using cert_pem or cert_file) and the client key (using key_pem or key_file) must be provided.

When min_version is not provided, the minimum acceptable TLS version is inherited from Go’s default minimum version, TLS 1.2. If min_version is provided, it must be set to one of the following strings:

• "TLS10" (TLS 1.0)
• "TLS11" (TLS 1.1)
• "TLS12" (TLS 1.2)
• "TLS13" (TLS 1.3)

queue_config block

Each queue then manages a number of concurrent shards which is responsible for sending a fraction of data to their respective endpoints. The number of shards is automatically raised if samples are not being sent to the endpoint quickly enough. The range of permitted shards can be configured with the min_shards and max_shards arguments.

Each shard has a buffer of samples it will keep in memory, controlled with the capacity argument. New metrics aren’t read from the WAL unless there is at least one shard that is not at maximum capacity.

The buffer of a shard is flushed and sent to the endpoint either after the shard reaches the number of samples specified by max_samples_per_send or the duration specified by batch_send_deadline has elapsed since the last flush for that shard.

Shards retry requests which fail due to a recoverable error. An error is recoverable if the server responds with an HTTP 5xx status code. The delay between retries can be customized with the min_backoff and max_backoff arguments.

The retry_on_http_429 argument specifies whether HTTP 429 status code responses should be treated as recoverable errors; other HTTP 4xx status code responses are never considered recoverable errors. When retry_on_http_429 is enabled, Retry-After response headers from the servers are honored.

metadata_config block

write_relabel_config block

The rule block contains the definition of any relabeling rules that can be applied to an input metric. If more than one rule block is defined, the transformations are applied in top-down order.

The following arguments can be used to configure a rule. All arguments are optional. Omitted fields take their default values.

You can use the following actions:

• drop - Drops metrics where regex matches the string extracted using the source_labels and separator.
• dropequal - Drop targets for which the concatenated source_labels do match target_label.
• hashmod - Hashes the concatenated labels, calculates its modulo modulus and writes the result to the target_label.
• keep - Keeps metrics where regex matches the string extracted using the source_labels and separator.
• keepequal - Drop targets for which the concatenated source_labels do not match target_label.
• labeldrop - Matches regex against all label names. Any labels that match are removed from the metric’s label set.
• labelkeep - Matches regex against all label names. Any labels that don’t match are removed from the metric’s label set.
• labelmap - Matches regex against all label names. Any labels that match are renamed according to the contents of the replacement field.
• lowercase - Sets target_label to the lowercase form of the concatenated source_labels.
• replace - Matches regex to the concatenated labels. If there’s a match, it replaces the content of the target_label using the contents of the replacement field.
• uppercase - Sets target_label to the uppercase form of the concatenated source_labels.

wal block

The wal block customizes the Write-Ahead Log (WAL) used to temporarily store metrics before they are sent to the configured set of endpoints.

The WAL serves two primary purposes:

• Buffer unsent metrics in case of intermittent network issues.
• Populate in-memory cache after a process restart.

The WAL is located inside a component-specific directory relative to the storage path Grafana Agent is configured to use. See the agent run documentation for how to change the storage path.

The truncate_frequency argument configures how often to clean up the WAL. Every time the truncate_frequency period elapses, the lower two-thirds of data is removed from the WAL and is no available for sending.

When a WAL clean-up starts, the lowest successfully sent timestamp is used to determine how much data is safe to remove from the WAL. The min_keepalive_time and max_keepalive_time control the permitted age range of data in the WAL; samples aren’t removed until they are at least as old as min_keepalive_time, and samples are forcibly removed if they are older than max_keepalive_time.

Exported fields

The following fields are exported and can be referenced by other components:

Component health

prometheus.remote_write is only reported as unhealthy if given an invalid configuration. In those cases, exported fields are kept at their last healthy values.

Debug information

prometheus.remote_write does not expose any component-specific debug information.

Debug metrics

• agent_wal_storage_active_series (gauge): Current number of active series being tracked by the WAL.
• agent_wal_storage_deleted_series (gauge): Current number of series marked for deletion from memory.
• agent_wal_out_of_order_samples_total (counter): Total number of out of order samples ingestion failed attempts.
• agent_wal_storage_created_series_total (counter): Total number of created series appended to the WAL.
• agent_wal_storage_removed_series_total (counter): Total number of series removed from the WAL.
• agent_wal_samples_appended_total (counter): Total number of samples appended to the WAL.
• agent_wal_exemplars_appended_total (counter): Total number of exemplars appended to the WAL.
• prometheus_remote_storage_samples_total (counter): Total number of samples sent to remote storage.
• prometheus_remote_storage_exemplars_total (counter): Total number of exemplars sent to remote storage.
• prometheus_remote_storage_metadata_total (counter): Total number of metadata entries sent to remote storage.
• prometheus_remote_storage_samples_failed_total (counter): Total number of samples that failed to send to remote storage due to non-recoverable errors.
• prometheus_remote_storage_exemplars_failed_total (counter): Total number of exemplars that failed to send to remote storage due to non-recoverable errors.
• prometheus_remote_storage_metadata_failed_total (counter): Total number of metadata entries that failed to send to remote storage due to non-recoverable errors.
• prometheus_remote_storage_samples_retries_total (counter): Total number of samples that failed to send to remote storage but were retried due to recoverable errors.
• prometheus_remote_storage_exemplars_retried_total (counter): Total number of exemplars that failed to send to remote storage but were retried due to recoverable errors.
• prometheus_remote_storage_metadata_retried_total (counter): Total number of metadata entries that failed to send to remote storage but were retried due to recoverable errors.
• prometheus_remote_storage_samples_dropped_total (counter): Total number of samples which were dropped after being read from the WAL before being sent to remote_write because of an unknown reference ID.
• prometheus_remote_storage_exemplars_dropped_total (counter): Total number of exemplars which were dropped after being read from the WAL before being sent to remote_write because of an unknown reference ID.
• prometheus_remote_storage_enqueue_retries_total (counter): Total number of times enqueue has failed because a shard’s queue was full.
• prometheus_remote_storage_sent_batch_duration_seconds (histogram): Duration of send calls to remote storage.
• prometheus_remote_storage_queue_highest_sent_timestamp_seconds (gauge): Unix timestamp of the latest WAL sample successfully sent by a queue.
• prometheus_remote_storage_samples_pending (gauge): The number of samples pending in shards to be sent to remote storage.
• prometheus_remote_storage_exemplars_pending (gauge): The number of exemplars pending in shards to be sent to remote storage.
• prometheus_remote_storage_shard_capacity (gauge): The capacity of shards within a given queue.
• prometheus_remote_storage_shards (gauge): The number of shards used for concurrent delivery of metrics to an endpoint.
• prometheus_remote_storage_shards_min (gauge): The minimum number of shards a queue is allowed to run.
• prometheus_remote_storage_shards_max (gauge): The maximum number of a shards a queue is allowed to run.
• prometheus_remote_storage_shards_desired (gauge): The number of shards a queue wants to run to be able to keep up with the amount of incoming metrics.
• prometheus_remote_storage_bytes_total (counter): Total number of bytes of data sent by queues after compression.
• prometheus_remote_storage_metadata_bytes_total (counter): Total number of bytes of metadata sent by queues after compression.
• prometheus_remote_storage_max_samples_per_send (gauge): The maximum number of samples each shard is allowed to send in a single request.
• prometheus_remote_storage_samples_in_total (counter): Samples read into remote storage.
• prometheus_remote_storage_exemplars_in_total (counter): Exemplars read into remote storage.

Examples

The following examples show you how to create prometheus.remote_write components that send metrics to different destinations.

Send metrics to a local Mimir instance

You can create a prometheus.remote_write component that sends your metrics to a local Mimir instance:

river Copy

prometheus.remote_write "staging" {
  // Send metrics to a locally running Mimir.
  endpoint {
    url = "http://mimir:9009/api/v1/push"

    basic_auth {
      username = "example-user"
      password = "example-password"
    }
  }
}

// Configure a prometheus.scrape component to send metrics to
// prometheus.remote_write component.
prometheus.scrape "demo" {
  targets = [
    // Collect metrics from Grafana Agent's default HTTP listen address.
    {"__address__" = "127.0.0.1:12345"},
  ]
  forward_to = [prometheus.remote_write.staging.receiver]
}

Send metrics to a Mimir instance with a tenant specified

You can create a prometheus.remote_write component that sends your metrics to a specific tenant within the Mimir instance. This is useful when your Mimir instance is using more than one tenant:

river Copy

prometheus.remote_write "staging" {
  // Send metrics to a Mimir instance
  endpoint {
    url = "http://mimir:9009/api/v1/push"

    headers = {
      "X-Scope-OrgID" = "staging",
    }
  }
}

Send metrics to a managed service

You can create a prometheus.remote_write component that sends your metrics to a managed service, for example, Grafana Cloud. The Prometheus username and the Grafana Cloud API Key are injected in this example through environment variables.

river Copy

prometheus.remote_write "default" {
  endpoint {
    url = "https://prometheus-xxx.grafana.net/api/prom/push"
      basic_auth {
        username = env("PROMETHEUS_USERNAME")
        password = env("GRAFANA_CLOUD_API_KEY")
      }
  }
}

Technical details

prometheus.remote_write uses snappy for compression.

Any labels that start with __ will be removed before sending to the endpoint.

Data retention

The prometheus.remote_write component uses a Write Ahead Log (WAL) to prevent data loss during network outages. The component buffers the received metrics in a WAL for each configured endpoint. The queue shards can use the WAL after the network outage is resolved and flush the buffered metrics to the endpoints.

The WAL records metrics in 128 MB files called segments. To avoid having a WAL that grows on-disk indefinitely, the component truncates its segments on a set interval.

On each truncation, the WAL deletes references to series that are no longer present and also checkpoints roughly the oldest two thirds of the segments (rounded down to the nearest integer) written to it since the last truncation period. A checkpoint means that the WAL only keeps track of the unique identifier for each existing metrics series, and can no longer use the samples for remote writing. If that data has not yet been pushed to the remote endpoint, it is lost.

This behavior dictates the data retention for the prometheus.remote_write component. It also means that it’s impossible to directly correlate data retention directly to the data age itself, as the truncation logic works on segments, not the samples themselves. This makes data retention less predictable when the component receives a non-consistent rate of data.

The WAL block in Flow mode, or the metrics config in Static mode contain some configurable parameters that can be used to control the tradeoff between memory usage, disk usage, and data retention.

The truncate_frequency or wal_truncate_frequency parameter configures the interval at which truncations happen. A lower value leads to reduced memory usage, but also provides less resiliency to long outages.

When a WAL clean-up starts, the most recently successfully sent timestamp is used to determine how much data is safe to remove from the WAL. The min_keepalive_time or min_wal_time controls the minimum age of samples considered for removal. No samples more recent than min_keepalive_time are removed. The max_keepalive_time or max_wal_time controls the maximum age of samples that can be kept in the WAL. Samples older than max_keepalive_time are forcibly removed.

Extended remote_write outages

When the remote write endpoint is unreachable over a period of time, the most recent successfully sent timestamp is not updated. The min_keepalive_time and max_keepalive_time arguments control the age range of data kept in the WAL.

If the remote write outage is longer than the max_keepalive_time parameter, then the WAL is truncated, and the oldest data is lost.

Intermittent remote_write outages

If the remote write endpoint is intermittently reachable, the most recent successfully sent timestamp is updated whenever the connection is successful. A successful connection updates the series’ comparison with min_keepalive_time and triggers a truncation on the next truncate_frequency interval which checkpoints two thirds of the segments (rounded down to the nearest integer) written since the previous truncation.

Falling behind

If the queue shards cannot flush data quickly enough to keep up-to-date with the most recent data buffered in the WAL, we say that the component is ‘falling behind’. It’s not unusual for the component to temporarily fall behind 2 or 3 scrape intervals. If the component falls behind more than one third of the data written since the last truncate interval, it is possible for the truncate loop to checkpoint data before being pushed to the remote_write endpoint.

WAL corruption

WAL corruption can occur when Grafana Agent unexpectedly stops while the latest WAL segments are still being written to disk. For example, the host computer has a general disk failure and crashes before you can stop Grafana Agent and other running services. When you restart Grafana Agent, the Agent verifies the WAL, removing any corrupt segments it finds. Sometimes, this repair is unsuccessful, and you must manually delete the corrupted WAL to continue.

If the WAL becomes corrupted, Grafana Agent writes error messages such as err="failed to find segment for index" to the log file.

To delete the corrupted WAL:

1. Stop Grafana Agent.

2. Find and delete the contents of the wal directory.

   By default the wal directory is a subdirectory of the data-agent directory located in the Grafana Agent working directory. The WAL data directory may be different than the default depending on the wal_directory setting in your Static configuration file or the path specified by the Flow command line flag --storage-path.

   Note

   There is one wal directory per:

   • Metrics instance running in Static mode
   • prometheus.remote_write component running in Flow mode

3. Start Grafana Agent and verify that the WAL is working correctly.