Event Correlation

opEvents does not just automatically suppress duplicate events (stateful or custom-matched); it can also create new events based on correlating recent event occurrences.

This event correlation and synthesis feature is configured in the same way as the duplicate suppression, namely by putting event creation rules into conf/EventRules.nmis.

An event creation rule consists of:

• an event name, which specifies the name of the newly created event,
• a list of events (more precisely, their names), which are the events to consider for correlation,
• a (minimum) count of events that have to be detected to trigger the rule,
• an optional list of groupby clauses, which define whether the count is interpreted globally for all named events, or separately within smaller groups,
• optional delayedaction and enrich clauses, which adjust the handling of the newly created event,
• and finally a window parameter, which defines the time window to examine.

(If you compare suppression and synthesis rules closely, you'll see that the main difference is the lack of a suppress clause for synthesis rules.)

Here is an example rule:

'3' => {
	name => 'Customer Outage',
 	events => ["Node Down","SNMP Down"],
 	window => '60',
 	count=> 5,
 	groupby=>['node.customer'], # count separately for every observed value of group
 	enrich=>{priority => 3, answer => 42}, # any such items gets inserted in the new event
},

This rule causes opEvents to look for Node Down and SNMP Down events in the last 60 seconds, separate them into per-customer groups (see grouping below); if it counts 5 or more such events in a group, then a new event called Customer Outage is created.

Event content and Enrichment

When a synthesis rule creates a new event, first the contents of the most recent triggering event are copied over. Then the event name is set to the name given in the rule, an audit trail of triggering events is added (by adding the properties nodes and eventids), the event is marked as synthetic, and finally any enrichment entries from the rule are added in.

Event Processing

At this point the new event is inserted into the database, and is ready for further action processing. This action processing (e.g. escalation, mail notification, custom logging) is performed either immediately or delayed after a custom delay period (if the delayedaction property is present in the rule).

Grouping

If no groupby clause is present, then the set of matching events is counted directly, which may be too generic for many common scenarios. For example creating new events for a particular customer or service group only wouldn't be possible. Grouping solves this problem: the set is split into groups with matching property values and the count threshold is applied to those groups.

The groupby clause has the form of a list of node.X or event.Y property specifications (e.g. node.customer or node.group), which are used to group events into buckets for counting: only events that share the same values for all the listed grouping properties will be counted together. For example, the groupby clause [ 'node.customer', 'event.priority' ] would cause this suppression rule to be applied independently for all combinations of customer and event priority. The clause given in the example block above will create a Customer Outage event for any individual customer with 5 outages in 60 seconds; without the groupby any 5 outages would cause a synthetic event to be created.

Synthetic events and Storm Control

All synthesis rules are applied independently, thus a single event could be a trigger for multiple synthetic events. This is desirable for example for detecting both per-customer problems and global issues at the same time: a few problem events can trigger a customer-specific action, while the same events could be counted together with others for detecting and reacting to a major outage.

However, great care has been taken to avoid event storms caused by synthetic events: When a synthesis rule fires because there were more than count matching events in the time window, then all the matching events are marked as consumed and will not be considered for any future synthesis for this rule. In other words, there is no overlap between successful synthesis time windows.

Here is a practical example for the consequences of this design: Let's assume a rule that specifies 5 event matches in 120 seconds as trigger. At some time T1 we count 25 such recent events, therefore the rule fires, a new event is created, and the 25 matches are consumed (not just the 5 that the trigger requires!). The count of triggering events thus starts from scratch at time T1. Let's assume that at T2, four seconds later, event correlation is performed next, and now only the events since T1 are considered as potential triggers. Assuming there were  3 bad events in these four seconds, no synthetic event will be created. Another 4 in the next few seconds, the count is now at 7 and the rule fires. On the other hand, if there had been 200 events between T1 and T2, then only one synthetic event would be created at time T2.

If a rule does not trigger because there are fewer than count trigger events, then naturally these events remain potential triggers until the time window moves past the events in question.