Seunghun Shin
512c292e04
What is this feature? This PR implements a jitter mechanism for periodic alert state storage to distribute database load over time instead of processing all alert instances simultaneously. When enabled via the state_periodic_save_jitter_enabled configuration option, the system spreads batch write operations across 85% of the save interval window, preventing database load spikes in high-cardinality alerting environments. Why do we need this feature? In production environments with high alert cardinality, the current periodic batch storage can cause database performance issues by processing all alert instances simultaneously at fixed intervals. Even when using periodic batch storage to improve performance, concentrating all database operations at a single point in time can overwhelm database resources, especially in resource-constrained environments. Rather than performing all INSERT operations at once during the periodic save, distributing these operations across the time window until the next save cycle can maintain more stable service operation within limited database resources. This approach prevents resource saturation by spreading the database load over the available time interval, allowing the system to operate more gracefully within existing resource constraints. For example, with 200,000 alert instances using a 5-minute interval and 4,000 batch size, instead of executing 50 batch operations simultaneously, the jitter mechanism distributes these operations across approximately 4.25 minutes (85% of 5 minutes), with each batch executed roughly every 5.2 seconds. This PR provides system-level protection against such load spikes by distributing operations across time, reducing peak resource usage while maintaining the benefits of periodic batch storage. The jitter mechanism is particularly valuable in resource-constrained environments where maintaining consistent database performance is more critical than precise timing of state updates.
Next generation alerting (ngalert) in Grafana 8
Ngalert (Next generation alert) is the next generation of alerting in Grafana 8.
Overview
The ngalert package can be found in pkg/services/ngalert and has the following sub-packages:
- api
- eval
- logging
- metrics
- models
- notifier
- schedule
- sender
- state
- store
- tests
Scheduling and evaluation of alert rules
The scheduling of alert rules happens in the schedule package. This package is responsible for managing the evaluation
of alert rules including checking for new alert rules and stopping the evaluation of deleted alert rules.
The scheduler runs at a fixed interval, called its heartbeat, in which it does a number of tasks:
- Fetch the alert rules for all organizations (excluding disabled)
- Start a goroutine (if this is a new alert rule or the scheduler has just started) to evaluate the alert rule
- Send an
*evaluationevent to the goroutine for each alert rule if its interval has elapsed - Stop the goroutines for all alert rules that have been deleted since the last heartbeat
The function that evaluates each alert rule is called run. It waits for an *evaluation event (sent each
interval seconds elapsed and is configurable per alert rule) and then evaluates the alert rule. To ensure that the
scheduler is evaluating the latest version of the alert rule it compares its local version of the alert rule with that
in the *evaluation event, fetching the latest version of the alert rule from the database if the version numbers
mismatch. It then invokes the Evaluator which evaluates any queries, classic conditions or expressions in alert rule
and passes the results of this evaluation to the State Manager. An evaluation can return no results in the case of
NoData or Error, a single result in the case of classic conditions, or more than one result if the alert rule is
multi-dimensional (i.e. one result per label set). In the case of multi-dimensional alert rules the results from an
evaluation should never contain more than one per label set.
The State Manager is responsible for determining the current state of the alert rule (normal, pending, firing, etc) by comparing each evaluation result to the previous evaluations of the same label set in the state cache. Given a label set, it updates the state cache with the new current state, the evaluation time of the current evaluation and appends the current evaluation to the slice of previous evaluations. If the alert changes state (i.e. pending to firing) then it also creates an annotation to mark it on the dashboard and panel for this alert rule.
You might have noticed that so far we have avoided using the word "Alert" and instead talked about evaluation results and the current state of an alert rule. The reason for that is at this time in the evaluation of an alert rule the State Manager does not know about alerts, it just knows for each label set the state of an alert rule, the current evaluation and previous evaluations.
Notification of alerts
When an evaluation transitions the state of an alert rule for a given label set from pending to firing or from firing to normal the scheduler creates an alert instance and passes it to Alertmanager. In the case where a label set is transitioning from pending to firing the state of the alert instance is "Firing" and when transitioning from firing to normal the state of the alert instance is "Normal".
Which Alertmanager?
In ngalert it is possible to send alerts to the internal Alertmanager, an external Alertmanager, or both.
The internal Alertmanager is called MultiOrgAlertmanager and creates an Alertmanager for each organization in
Grafana to preserve isolation between organizations in Grafana. The MultiOrgAlertmanager receives alerts from the
scheduler and then forwards the alert to the correct Alertmanager for the organization.
When Grafana is configured to send alerts to an external Alertmanager it does so via the sender which creates an
abstraction over notification of alerts and discovery of external Alertmanagers in Prometheus. The sender receives
alerts via the SendAlerts function and then passes them to Prometheus.
How does Alertmanager turn alerts into notifications?
Alertmanager receives alerts via the PutAlerts function. Each alert is validated and its annotations and labels are
normalized, then the alerts are put in an in-memory structure. The dispatcher iterates over the alerts and matches
it to a route in the configuration as explained here.
The alert is then matched to an alert group depending on the configuration in the route. The alert is then sent through a number of stages including silencing and inhibition (which is currently not supported) and at last the receiver which can include wait, de-duplication, retry.
What are notification channels?
Notification channels receive alerts and turn them into notifications and is often the last callback in the receiver after wait, de-duplication and retry.