Scraping Service Mode
Scraping Service Mode is a third operational mode of the Grafana Agent that allows for clustering a set of Agent processes and distributing scrape load across them.
Determining what to scrape is done by writing instance configuration files to an API, which then stores the configuration files in a KV store backend. All agents in the cluster must use the same KV store so they see the same set of config files.
Each process of the Grafana Agent can be running multiple independent “instances” at once, where an “instance” refers to the combination of:
- Service discovery for all
scrape_configswithin that loaded config
- Scrapes metrics from all discovered targets
- Stores data in its own Write-Ahead Log specific to the loaded config
- Remote Writes scraped metrics to the configured
remote_writedestinations specified within the loaded config.
The “instance configuration file,” then, is the configuration file that
specifies the set of
remote_write endpoints. For example,
a small instance configuration file looks like:
scrape_configs: - job_name: self-scrape static_configs: - targets: ['localhost:9090'] labels: process: 'agent' remote_write: - url: http://cortex:9009/api/prom/push
The full set of supported options for an instance configuration file is
available in the
Having multiple instance configuration files is necessary for sharding; each config file is distributed to a particular agent on the cluster based on the hash of its contents.
When Scraping Service Mode is enabled, Agents disallow specifying
instance configurations locally in the configuration file; using the KV store
agentctl can be used to manually sync
instance configuration files to the Agent’s API server.
Distributed hash ring
Scraping Service Mode uses a Distributed Hash Ring (commonly just called “the ring”) to cluster agents and to shard configurations within that ring. Each Agent joins the ring with a random distinct set of tokens that are used for sharding. The default number of generated tokens is 128.
The Distributed Hash Ring is also stored in a KV store. Since a KV store is also needed for storing configuration files, it is encouraged to re-use the same KV store for the ring.
When sharding, the Agent currently uses the name of a config file stored in the KV store for load distribution. Config names are guaranteed to be unique keys. The hash of the name is used as the lookup key in the ring and determines which agent (based on token) should be responsible for that config. “Price is Right” rules are used for the Agent lookup; the Agent owning the token with the closest value to the key without going over is responsible for the config.
All Agents are simultaneously watching the KV store for changes to the set of configuration files. When a config file is added or updated in the configuration store, each Agent will run the config name hash through their copy of the Hash Ring to determine if they are responsible for that config.
When an Agent receives a new config that it is responsible for, it launches a new instance from the instance config. If a config is deleted from the KV store, this will be detected by the owning Agent and it will stop the metric collection process for that config file.
When an Agent receives an event for an updated configuration file that they used to be the owner of but are no longer the owner, the associated instance for that configuration file is stopped for that Agent. This can happen when the cluster size changes.
Scraping Service Mode currently does not support replication; only one agent at a time will be responsible for scraping a certain config.
When a new Agent joins or leaves the cluster, the set of tokens in the ring may cause configurations to hash to a new Agent. The process of responding to this action is called “resharding.”
Resharding is run:
- When an Agent joins the ring
- When an Agent leaves the ring
- When the KV store sends a notification indicating a config has changed.
- On a specified interval in case KV change events have not fired.
The resharding process involves each Agent retrieving the full set of configurations stored in the KV store and determining if:
- The config owned by the current resharding Agent has changed and needs to be reloaded.
- The config is no longer owned by the current resharding Agent and the associated instance should be stopped.
- The config has been deleted and the associated instance should be stopped.
Because distribution is determined by the number of config files and not how many targets exist per config file, the best amount of distribution is achieved when each config file has the lowest amount of targets possible. The best distribution will be achieved if each config file stored in the KV store is limited to one static config with only one target.
A better distribution mechanism that distributes based on discovered targets is planned for the future.
Here’s an example
agent.yaml config file that uses the same
etcd server for
both configuration storage and the distributed hash ring storage:
server: log_level: debug http_listen_port: 12345 metrics: global: scrape_interval: 1m scraping_service: enabled: true kvstore: store: etcd etcd: endpoints: - etcd:2379 lifecycler: ring: replication_factor: 1 kvstore: store: etcd etcd: endpoints: - etcd:2379
Note that there are no instance configs present in this example; instance configs must be passed to the API for the Agent to start scraping metrics. See the docker-compose Scraping Service Example for how to run a Scraping Service Agent cluster locally.
agentctl is a tool included with this repository that helps users interact
with the new Config Management API. The
agentctl config-sync subcommand uses
local YAML files as a source of truth and syncs their contents with the API.
Entries in the API not in the synced directory will be deleted.
agentctl is distributed in binary form with each release and as a Docker
container with the
grafana/agentctl image. Tanka configurations that
grafana/agentctl and sync a set of configurations to the API
are planned for the future.