Alarm

Alarm core is driven by a collection of rules, which are defined in config/alarm-settings.yml. There are three parts in alarm rule definition.

1. Alarm rules. They define how metrics alarm should be triggered, what conditions should be considered.
2. Webhooks. The list of web service endpoint, which should be called after the alarm is triggered.
3. gRPCHook. The host and port of remote gRPC method, which should be called after the alarm is triggered.

Rules

Alarm rule is constituted by following keys

• Rule name. Unique name, show in alarm message. Must end with _rule.
• Metrics name. A.K.A. metrics name in oal script. Only long, double, int types are supported. See List of all potential metrics name.
• Include names. The following entity names are included in this rule. Such as Service name, endpoint name.
• Exclude names. The following entity names are excluded in this rule. Such as Service name, endpoint name.
• Threshold. The target value. For multiple values metrics, such as percentile, the threshold is an array. Described like value1, value2, value3, value4, value5. Each value could the threshold for each value of the metrics. Set the value to - if don't want to trigger alarm by this or some of the values.
  Such as in percentile, value1 is threshold of P50, and -, -, value3, value4, value5 means, there is no threshold for P50 and P75 in percentile alarm rule.
• OP. Operator, support >, >=, <, <=, =. Welcome to contribute all OPs.
• Period. How long should the alarm rule should be checked. This is a time window, which goes with the backend deployment env time.
• Count. In the period window, if the number of values over threshold(by OP), reaches count, alarm should send.
• Silence period. After alarm is triggered in Time-N, then keep silence in the TN -> TN + period. By default, it is as same as Period, which means in a period, same alarm(same ID in same metrics name) will be trigger once.

rules:
  # Rule unique name, must be ended with `_rule`.
  endpoint_percent_rule:
    # Metrics value need to be long, double or int
    metrics-name: endpoint_percent
    threshold: 75
    op: <
    # The length of time to evaluate the metrics
    period: 10
    # How many times after the metrics match the condition, will trigger alarm
    count: 3
    # How many times of checks, the alarm keeps silence after alarm triggered, default as same as period.
    silence-period: 10
  service_percent_rule:
    metrics-name: service_percent
    # [Optional] Default, match all services in this metrics
    include-names:
      - service_a
      - service_b
    exclude-names:
      - service_c
    # Single value metrics threshold.
    threshold: 85
    op: <
    period: 10
    count: 4
  service_resp_time_percentile_rule:
    # Metrics value need to be long, double or int
    metrics-name: service_percentile
    op: ">"
    # Multiple value metrics threshold. Thresholds for P50, P75, P90, P95, P99.
    threshold: 1000,1000,1000,1000,1000
    period: 10
    count: 3
    silence-period: 5
    message: Percentile response time of service {name} alarm in 3 minutes of last 10 minutes, due to more than one condition of p50 > 1000, p75 > 1000, p90 > 1000, p95 > 1000, p99 > 1000

Default alarm rules

We provided a default alarm-setting.yml in our distribution only for convenience, which including following rules

1. Service average response time over 1s in last 3 minutes.
2. Service success rate lower than 80% in last 2 minutes.
3. Percentile of service response time is over 1s in last 3 minutes
4. Service Instance average response time over 1s in last 2 minutes.
5. Endpoint average response time over 1s in last 2 minutes.

List of all potential metrics name

The metrics names are defined in official OAL scripts, right now metrics from Service, Service Instance, Endpoint scopes could be used in Alarm, we will extend in further versions.

Submit issue or pull request if you want to support any other scope in alarm.

Webhook

Webhook requires the peer is a web container. The alarm message will send through HTTP post by application/json content type. The JSON format is based on List<org.apache.skywalking.oap.server.core.alarm.AlarmMessage with following key information.

• scopeId, scope. All scopes are defined in org.apache.skywalking.oap.server.core.source.DefaultScopeDefine.
• name. Target scope entity name.
• id0. The ID of scope entity, matched the name.
• id1. Not used today.
• ruleName. The rule name you configured in alarm-settings.yml.
• alarmMessage. Alarm text message.
• startTime. Alarm time measured in milliseconds, between the current time and midnight, January 1, 1970 UTC.

Example as following

[{
	"scopeId": 1, 
        "scope": "SERVICE",
        "name": "serviceA", 
	"id0": "12",  
	"id1": "",  
        "ruleName": "service_resp_time_rule",
	"alarmMessage": "alarmMessage xxxx",
	"startTime": 1560524171000
}, {
	"scopeId": 1,
        "scope": "SERVICE",
        "name": "serviceB",
	"id0": "23",
	"id1": "",
        "ruleName": "service_resp_time_rule",
	"alarmMessage": "alarmMessage yyy",
	"startTime": 1560524171000
}]

gRPCHook

The alarm message will send through remote gRPC method by Protobuf content type. The message format with following key information which are defined in oap-server/server-alarm-plugin/src/main/proto/alarm-hook.proto.

Part of protocol looks as following:

message AlarmMessage {
    int64 scopeId = 1;
    string scope = 2;
    string name = 3;
    string id0 = 4;
    string id1 = 5;
    string ruleName = 6;
    string alarmMessage = 7;
    int64 startTime = 8;
}

Update the settings dynamically

Since 6.5.0, the alarm settings can be updated dynamically at runtime by Dynamic Configuration, which will override the settings in alarm-settings.yml.

In order to determine that whether an alarm rule is triggered or not, SkyWalking needs to cache the metrics of a time window for each alarm rule, if any attribute (metrics-name, op, threshold, period, count, etc.) of a rule is changed, the sliding window will be destroyed and re-created, causing the alarm of this specific rule to restart again.