9.9 KiB
Vitrage Evaluator Design
Overview
The Vitrage Scenario Evaluator is charged with the evaluation of the scenarios specified in the Vitrage templates, and executing the actions associated with these scenarios when they are triggered. The Scenario Evaluator is also responsible for resolving any inter-action dependencies, such as two contradicting actions (e.g., raise and disable the same alarm), two overlapping actions etc.
Evaluator Design
+-------------+
| +-------------+
| | |
| | Templates |
+-+ (YAML) |
+-----------+-+
|
+------------------------------------------------------------------------------+
|Template | |
|Loading | load |
| | |
| +-----v-------------+ +---------------+ +---------------+ |
| | Template | | | | | |
| | |process | Template | add | Scenario | |
| |(Python Dictionary)+--------> Object +-------> Repo | |
| | | | | | | |
| +-------------------+ +---------------+ +---+--------+--+ |
| | ^ |
| | | |
| | | |
+------------------------------------------------------------------------------+
| |
+-----------------------------------------------------------------------------------+
| Event Processing return relevant| | get |
| scenarios | | scenarios |
| +--------------+ +----------------------------------+ |
| | | perform | analyze | | | |
| | Action | action | matches | | | |
| | Executor <--------------------------------+ | | | |
| | | | | | | | |
| | | | | | | | |
| +--------------+ +----------------------------------+ |
| | | | |
| return scenario | | | |
| matches | | | |
| +-----+------v--------+ |
| | | |
| | Entity Graph | |
| | | |
| | | |
| +---------------------+ |
+-----------------------------------------------------------------------------------+Flow
Concepts and Guidelines
- Events: The Scenario Evaluator is notified of each event in the Entity Graph after the event takes place. An event in this context is any change (create/update/delete) in a graph element (vertex/edge). The notification will consist of two copies of the element: the element before the change and the current element after the change took place.
- Actions - Do and Undo: If the Entity Graph matches a scenario, the relevant actions will need to be executed. Conversely, if a previously matched scenario no longer holds, the action needs to be undone. Thus, for each action there must be a "do" and "undo" procedure defined. For example, the raise_alarm action will raise an alarm in the "do" procedure, and disable the alarm in the "undo" procedure. An important feature of this approach is that for a given scenario and a specific event, either the "do" or "undo" phases can be performed but not both simultaneously.
Template Loading
Scenarios are written up in configuration files called templates. The format and specification of these can be seen here.
Templates should all be located in the <vitrage folder>/templates folder.
When Vitrage is started up, all the templates are loaded into a Scenario Repository. During this loading, template verification is done to ensure the correct format is used, references are valid, and more. Errors in each template should be written to the log. Invalid templates are skipped.
The Scenario Repository supports querying for scenarios based on a vertex or edge in the Entity Graph. Given such a graph element, the Scenario Repository will return all scenarios where this element appears in the scenario condition. This means that a corresponding element appears in the scenario condition, such that for each key-value in the template, the same key-value can be found in the element (but not always the reverse).
Ongoing Operation
- The Scenario Evaluator is notified of an event on some element in the Entity Graph.
- The Scenario Evaluator queries the Scenario Repository for relevant scenarios for both before and current states of the element, and returns a set of matching scenarios for each.
- The two sets of scenarios are analyzed and filtered, resulting in two disjoint sets, to avoid "do/undo" conflicts (See above in 'Concepts and Guidelines'_). Currently, this filtering will be done by removing any scenario that appears in both from both sets.
- For each scenario related to the before element, the Scenario Evaluator queries the Entity Graph for all the matching patterns in the current system. For each match and each associated action, a reference to the undo of the action is added to an action collection.
- For each scenario related to the current element,the Scenario Evaluator queries the Entity Graph for all the matching patterns in the current system. For each match and each associated action, a reference to the do of the action is added to the same action collection.
- Given all the actions (do & undo) in the action collection, analyze the dependencies between themselves, as well as between them and currently active actions. The result of this analysis will be a new collection of actions comprised of the correct actions to perform. See below in the 'Dealing with Overlapping Scenarios'_ Section.
- Given all the actions (do & undo) in the action
collection, perform them using action executor.
- Currently, the only action filtering is avoiding performing the same action twice.
System Initialization
During the initialization of Vitrage, the Scenario Evaluator will be de-activated until all the datasources complete their initial "get_all" process. After it is activated, the consistency flow will begin, which will trigger all the relevant scenarios for each element in the Entity Graph.
This approach has several benefits:
- During the initialization period, many events need to be processed into the Entity Graph. By postponing the evaluation till after this period, we avoid bottlenecks and other performance issues.
- During the initialization period the Entity Graph is built step-by-step until it reflects the current status of the Cloud. Thus, during this interim period scenarios that contain "not" clauses might be triggered because a certain entity is not present in the graph, even though it is present in reality and just has not been processed into the graph (since the "get_all" is not finished).
It is possible that this late activation of the evaluator will be removed or changed once we move to a persistent graph database for the Entity Graph in future version.
Dealing with Overlapping Scenarios
There can be multiple Vitrage scenarios loaded in a specific system, some of which might overlap in their targets. For example, two scenarios might have a "set_state" action, with identical or different states, for the same resource. We need to deal with such overlaps. Currently, the goal is to support overlap of the same action type with itself, specifically the following use cases, which correspond to the three actions Vitrage supports at this point:
- set_state: Two scenarios setting the state of the same resource
- raise_alarm: Two scenarios raising the same deduced alarm (with possibly different severity)
- add_causal_relationship: Two scenarios adding the same causal relationship
For all of these, the desired behavior is to choose the dominant outcome or action. For set_state this means the worst state, and for raise_alarm this means the highest severity of all active actions.
In order to support this feature, Vitrage maintains an in-memory record of all active actions, indexed (broadly) according to their affected target. This allows for tracking individual actions and their triggers, even when they overlap in their effect.
For more details on the implementation of this functionality, see the design on this etherpad.