opEvents does not just automatically suppress duplicate events (stateful or custom-matched); it can also create new events based on correlating recent event occurrences.
In versions 2.0.4 and newer you can use fine-grained controls to deal with the triggering events, and from version 2.2 onwards the contents of synthetic events are configurable, too.
This page describes how to configure event correlation.
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 synthesis rule consists of:
name
, which specifies the name of the newly created event,events
(more precisely, their names), which are the events to consider for correlation,count
of events that have to be detected to trigger the rule,groupby
clauses, which define whether the count is interpreted globally for all named events, or separately within smaller groups,delayedaction
and autoacknowledge
clauses, which define how the triggering events should be handled,enrich
clause, which adjusts the content of the newly created event,copy_first
, copy_last
, copy_highest
and copy_groupby
clauses which further control the contents of the newly created event,inhibit
parameter, which disables correlation temporarily after a rule has fired,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, whereas the suppression rules don't have enrich
or copy_*
clauses.)
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 customer 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.
If no groupby
clause is present, then potential trigger events are selected solely by event name and event time (within the window), without any further scope limiting, i.e. globally across all nodes. For many common scenarios this may be too broad a selection; for example creating new events for a particular customer or service group only wouldn't be possible.
Grouping solves this problem: the set of potential triggering events 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 correlation 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 anywhere would cause a synthetic event to be created.
The groupby
clause can make use of all common node properties which are listed here and the standard event properties which are documented on this page. Please note, however, that only event properties that were set during the event parsing stage are accessible when correlation is performed. For example policy actions can change an event (e.g. tagging, script execution) but policy actions are performed after correlation.
Before opEvents version 2.2, synthetic events are always cloned from the most recent triggering event, then they get a new name from the synthesis rule name, and finally any static enrich
clauses are evaluated. Synthetic events could not be stateful events, i.e. they were not subject to deduplication and could not be acknowledged (or 'closed') by any later 'opposite' event.
In version 2.2 this limitation has been removed, and much more precise control of the event content is possible.
When a synthesis rule creates a new event, the following steps are performed:
copy_first
, copy_last
, copy_highest, copy_present
or copy_groupby
directives are present, then a backwards-compatible directive 'copy_last => [ qr// ]'
is added.copy_present
is evaluated first. It specifies which properties should be set from their first occurrence.copy_present => [ 'alpha', 'beta' ]
will pull the alpha
and beta
properties from wherever they are present for the first time, but independent of each other: a trigger event can contribute none, either or both properties.copy_first
is evaluated next, and specifies which event properties should be copied over from the earliest trigger event.qr/cust.*/
, then all properties with names matching the regular expression are copied.copy_last
is checked next, and properties listed here are copied over from the most recent trigger event.copy_first
).copy_highest
is checked next, and its properties are sourced from the trigger event with the highest priority.copy_groupby
controls whether any of the grouping property values should be saved in the new event.groupby
directive.groupby
list, the value of the grouping property is saved as the target name in the new event, if a target name is available in the copy_groupby
list.groupby
is given for this rule, then a copy_groupby
directive has no effect.enrich
clause is checked, and each of its property name - value pairs indicates which properties should be set to (or overwritten with) a particular static value.node
, stateful
and element
properties are automatically adjusted if required (see below for details).nodes
and eventids
), the event is marked as synthetic
and is inserted into the database.Please note that "earliest event" in step 2, 3 and 4 refers to the event with the earliest event timestamp, which does not necessarily reflect its processing order. opEvents processes inputs mostly - but not always - in chronological order. If you have multiple 'earliest' events (all with the same timestamp) then their order is undefined and copy_first
will pick a random event. The same caveat applies for the "most recent event".
If no copy_*
or enrich
clause has caused the the node
property to be set explicitly, then the global default node is used instead.
'Set explicitly' means a copy_*
or enrich
clause did include the node
property, i.e. not if the node property copying happened because of a regular expression.
The global default node in opEvents 2.2 is configurable using the configuration item opevents_correlation_node
, and it's normally called "global".
This virtual node is automatically (re)created if missing.
This behaviour is different from opEvents before 2.2, where all synthetic events were attached to the last trigger event's node. To emulate the old behaviour you have to change your correlation rules, so that they include the directive
copy_last => [qr//, 'node']
which causes a blanket copy of all properties from the last trigger event and an explicit copy of the node property (to disable the automatic event node choice).
Here is an example rule demonstrating the new directives:
'1' => { name => "Very Sick Node", events => [ "Node Down", "SNMP Down", "Interface Down", "Service Down", "Service Degraded", "Interface Flap", "Node Flap", "WMI Down" ], window => 120, count => 3, groupby => [ 'node.name' ], # we want separate events for each node of course enrich => { stateful => "Very Sick Node", priority => 5, state => 'down', element => undef }, # new event is stateful only if stateful is set or copied by name copy_last => [ qr//, 'node' ], # can set from node here (all events share it) copy_groupby => [ 'node' ], # or from here; must set it explicitely somewhere, or the event goes to opevents_correlation_node }, |
By default, synthetic events are not stateful events, i.e. they are not subject to deduplication and they cannot be acknowledged (or 'closed') by any future 'opposite' event.
However, in 2.2 and newer it is possible to enable stateful handling for synthetic events:
stateful
property. copy_*
does not meet this requirement, and a thusly copied stateful
property is deleted before event creation.state
property value is present.element
property value is present, or opEvents will automatically create one from groupby
information if that is available.node
, stateful
and element
properties must uniquely identify the stateful 'thing', and the value of the state
property describes the new state.The example rule above shows how a stateful 'very sick node' event can be created: the node name is set from the grouping criteria (i.e. all related triggers share the same node name), the stateful
property is set with a static enrich
clause, and there is no element
, so at most one 'very sick node' stateful thing can exist for a single node.
If we wanted to acknowledge this event from a different correlation rule, we'd have to ensure that node
, stateful
and element
properties with the same value are generated, but the state
would have to be 'up' or 'closed' or 'ok'.
Here is another example, for a group-level stateful event:
8 => { name=>'sick group', events=>["Service Down","SNMP Down", "Node Down"], groupby => [ 'node.group' ], window=>150, count=>3, enrich => { stateful => "sick group", state => "down" }, # node will be opevents_correlation_node, element will be group copy_last => [ qr// ], }, 9 => { name=>'happy group', events=>["Service Up", "some nice event" ], groupby => [ 'node.group' ], window=>300, count=>1, enrich => { stateful => "sick group", state => "up" }, copy_last => [ qr// ], }, |
Rule 8 specifies that three of the listed 'down' events in a single group should cause a new event that sets the 'sick group' state to down for this one group; the element
property is auto-generated from the groupby
data, and all such events are attached to the virtual node 'global'. Any repeat 'sick group' events would be statefully deduplicated. Because element is set to the group in question, every single group would have its own 'sick group' state.
As soon as a single positive event from the list in rule 9 arrives, the 'sick group' event is acknowledged and closed.
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 immediately.
Please note that this feature is only available in opEvents 2.0.4 and newer.
delayedaction
clause (with a numeric value), then all potentially triggering events will have their action processing delayed by the given number of seconds.delayaction
value should therefore be set to a relatively small value.autoacknowledge
set to "true"
or 1, then all triggering events will be automatically acknowledged and all action processing for them will be aborted.The combination of these two controlling properties provides fine-grained storm control and the ability to create "combination events" that subsume and close any number of triggering events:
If your rule sets neither delayedaction
nor autoacknowledge
, then the incoming potential trigger events will be processed as per normal and immediately, and any policy actions for them will be taken as soon as possible (but possible after being delayed by the state_flap_window
- see Deduplication and storm control in opEvents for details). The trigger events are thus completed, and visible as current/unacknowledged, completely independent of any synthetic events that might get triggered by them later.
If your rule sets delayedaction
(and optionally autoacknowledge
), then the incoming events are delayed and held for the given time before any policy actions are taken for them. (The delayedaction
setting should be the same as or larger than the rule's window
setting.)
If the requirements for a synthetic event are met during that time, then the new synthetic event can "combine" and supersede the triggering events. If autoacknowledge
is set, then all the triggering events will be acknowledged, closed and no actions will be taken for them at all. (Without autoacknowledge
the triggering events would not have actions performed but would remain unacknowledged.)
The net effect is that the current events view would show only the new synthetic event as 'current' and all the underlying triggering events would be categorized as closed (and optionally acknowledged), and thus be mostly hidden.
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.
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. 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 them.
However, synthetic event creation currently happens immediately as soon as a sufficient number of triggers are detected: assuming a trigger of a minimum 20 events in 60 seconds, receiving 100 events in that time frame will cause a new synthetic event for each of the 20 sufficient triggers.
Version 2.2 provides a new capability for fine-tuning storm control: the inhibit
timer.
If a correlation rule fires, and if that rule contains a numeric inhibit
parameter greater than zero, then opEvents will temporarily disable the rule with its particular groupby
context for that many seconds.
The primary application of this feature is to stop 'nuisance' repeat synthetics if a very large number of triggers arrives in a very short time frame: it lets you tell opEvents to generate at most one instance of a particular event every inhibit
seconds.
Here is an example scenario: let's assume a rule for raising a 'Group Outage' event if 20 instances of a particular event are seen within a window of 60 seconds. A major outage happens, and 100 such trigger events for group A arrive within just a seconds, and a further 25 triggers for group B.
inhibit
, after the first 20 events for group A you'll get one synthetic event for group A; another after the next 20 and so on. inhibit
set to 40 seconds (for example), you'll get the very first group A synthetic event as before, but then no synthetic events for this rule and group A for the next 40s;groupby
context.