Log Analysis Language

Log Analysis Language (LAL) in SkyWalking is essentially a Domain-Specific Language (DSL) to analyze logs. You can use LAL to parse, extract, and save the logs, as well as collaborate the logs with traces (by extracting the trace ID, segment ID and span ID) and metrics (by generating metrics from the logs and sending them to the meter system).

The LAL config files are in YAML format, and are located under directory lal. You can set log-analyzer/default/lalFiles in the application.yml file or set environment variable SW_LOG_LAL_FILES to activate specific LAL config files.

Filter

A filter is a group of parser, extractor and sink. Users can use one or more filters to organize their processing logic. Every piece of log will be sent to all filters in an LAL rule. A piece of log sent to the filter is available as property log in the LAL, therefore you can access the log service name via log.service. For all available fields of log, please refer to the protocol definition.

All components are executed sequentially in the orders they are declared.

Global Functions

Globally available functions may be used them in all components (i.e. parsers, extractors, and sinks) where necessary.

  • abort

By default, all components declared are executed no matter what flags (dropped, saved, etc.) have been set. There are cases where you may want the filter chain to stop earlier when specified conditions are met. abort function aborts the remaining filter chain from where it’s declared, and all the remaining components won’t be executed at all. abort function serves as a fast-fail mechanism in LAL.

filter {
    if (log.service == "TestingService") { // Don't waste resources on TestingServices
        abort {} // all remaining components won't be executed at all
    }
    // ... parsers, extractors, sinks
}

Note that when you put regexp in an if statement, you need to surround the expression with () like regexp(<the expression>), instead of regexp <the expression>.

Parser

Parsers are responsible for parsing the raw logs into structured data in SkyWalking for further processing. There are 3 types of parsers at the moment, namely json, yaml, and text.

When a piece of log is parsed, there is a corresponding property available, called parsed, injected by LAL. Property parsed is typically a map, containing all the fields parsed from the raw logs. For example, if the parser is json / yaml, parsed is a map containing all the key-values in the json / yaml; if the parser is text , parsed is a map containing all the captured groups and their values (for regexp and grok).

All parsers share the following options:

Option Type Description Default Value
abortOnFailure boolean Whether the filter chain should abort if the parser failed to parse / match the logs true

See examples below.

json

filter {
    json {
        abortOnFailure true // this is optional because it's default behaviour
    }
}

yaml

filter {
    yaml {
        abortOnFailure true // this is optional because it's default behaviour
    }
}

text

For unstructured logs, there are some text parsers for use.

  • regexp

regexp parser uses a regular expression (regexp) to parse the logs. It leverages the captured groups of the regexp, all the captured groups can be used later in the extractors or sinks. regexp returns a boolean indicating whether the log matches the pattern or not.

filter {
    text {
        abortOnFailure true // this is optional because it's default behaviour
        // this is just a demo pattern
        regexp "(?<timestamp>\\d{8}) (?<thread>\\w+) (?<level>\\w+) (?<traceId>\\w+) (?<msg>.+)"
    }
    extractor {
        tag level: parsed.level
        // we add a tag called `level` and its value is parsed.level, captured from the regexp above
        traceId parsed.traceId
        // we also extract the trace id from the parsed result, which will be used to associate the log with the trace
    }
    // ...
}
  • grok (TODO)

We’re aware of certains performance issues in the grok Java library, and so we’re currently conducting investigations and benchmarking. Contributions are welcome.

Extractor

Extractors aim to extract metadata from the logs. The metadata can be a service name, a service instance name, an endpoint name, or even a trace ID, all of which can be associated with the existing traces and metrics.

  • service

service extracts the service name from the parsed result, and set it into the LogData, which will be persisted (if not dropped) and is used to associate with traces / metrics.

  • instance

instance extracts the service instance name from the parsed result, and set it into the LogData, which will be persisted (if not dropped) and is used to associate with traces / metrics.

  • endpoint

endpoint extracts the service instance name from the parsed result, and set it into the LogData, which will be persisted (if not dropped) and is used to associate with traces / metrics.

  • traceId

traceId extracts the trace ID from the parsed result, and set it into the LogData, which will be persisted (if not dropped) and is used to associate with traces / metrics.

  • segmentId

segmentId extracts the segment ID from the parsed result, and set it into the LogData, which will be persisted (if not dropped) and is used to associate with traces / metrics.

  • spanId

spanId extracts the span ID from the parsed result, and set it into the LogData, which will be persisted (if not dropped) and is used to associate with traces / metrics.

  • timestamp

timestamp extracts the timestamp from the parsed result, and set it into the LogData, which will be persisted (if not dropped) and is used to associate with traces / metrics.

The unit of timestamp is millisecond.

  • layer

layer extracts the layer from the parsed result, and set it into the LogData, which will be persisted (if not dropped) and is used to associate with service.

  • tag

tag extracts the tags from the parsed result, and set them into the LogData. The form of this extractor should look something like this: tag key1: value, key2: value2. You may use the properties of parsed as both keys and values.

import javax.swing.text.LayeredHighlighter

filter {
    // ... parser

    extractor {
        tag level: parsed.level, (parsed.statusCode): parsed.statusMsg
        tag anotherKey: "anotherConstantValue"
        layer 'GENERAL'
    }
}
  • metrics

metrics extracts / generates metrics from the logs, and sends the generated metrics to the meter system. You may configure MAL for further analysis of these metrics. The dedicated MAL config files are under directory log-mal-rules, and you can set log-analyzer/default/malFiles to enable configured files.

# application.yml
# ...
log-analyzer:
  selector: ${SW_LOG_ANALYZER:default}
  default:
    lalFiles: ${SW_LOG_LAL_FILES:my-lal-config} # files are under "lal" directory
    malFiles: ${SW_LOG_MAL_FILES:my-lal-mal-config,another-lal-mal-config} # files are under "log-mal-rules" directory

Examples are as follows:

filter {
    // ...
    extractor {
        service parsed.serviceName
        metrics {
            name "log_count"
            timestamp parsed.timestamp
            labels level: parsed.level, service: parsed.service, instance: parsed.instance
            value 1
        }
        metrics {
            name "http_response_time"
            timestamp parsed.timestamp
            labels status_code: parsed.statusCode, service: parsed.service, instance: parsed.instance
            value parsed.duration
        }
    }
    // ...
}

The extractor above generates a metrics named log_count, with tag key level and value 1. After that, you can configure MAL rules to calculate the log count grouping by logging level like this:

# ... other configurations of MAL

metrics:
  - name: log_count_debug
    exp: log_count.tagEqual('level', 'DEBUG').sum(['service', 'instance']).increase('PT1M')
  - name: log_count_error
    exp: log_count.tagEqual('level', 'ERROR').sum(['service', 'instance']).increase('PT1M')

The other metrics generated is http_response_time, so you can configure MAL rules to generate more useful metrics like percentiles.

# ... other configurations of MAL

metrics:
  - name: response_time_percentile
    exp: http_response_time.sum(['le', 'service', 'instance']).increase('PT5M').histogram().histogram_percentile([50,70,90,99])

Sink

Sinks are the persistent layer of the LAL. By default, all the logs of each filter are persisted into the storage. However, some mechanisms allow you to selectively save some logs, or even drop all the logs after you’ve extracted useful information, such as metrics.

Sampler

Sampler allows you to save the logs in a sampling manner. Currently, the following sampling strategies are supported:

  • rateLimit: samples n logs at a maximum rate of 1 minute. rateLimit("SamplerID") requires an ID for the sampler. Sampler declarations with the same ID share the same sampler instance, thus sharing the same rpm and resetting logic.
  • possibility: every piece of log has a pseudo possibility of percentage to be sampled, the possibility was generated by Java random number generator and compare to the given percentage option.

We welcome contributions on more sampling strategies. If multiple samplers are specified, the last one determines the final sampling result. See examples in Enforcer.

Examples 1, rateLimit:

filter {
    // ... parser

    sink {
        sampler {
            if (parsed.service == "ImportantApp") {
                rateLimit("ImportantAppSampler") {
                    rpm 1800  // samples 1800 pieces of logs every minute for service "ImportantApp"
                }
            } else {
                rateLimit("OtherSampler") {
                    rpm 180   // samples 180 pieces of logs every minute for other services than "ImportantApp"
                }
            }
        }
    }
}

Examples 2, possibility:

filter {
    // ... parser

    sink {
        sampler {
            if (parsed.service == "ImportantApp") {
                possibility(80) { // samples 80% of the logs for service "ImportantApp"
                }
            } else {
                possibility(30) { // samples 30% of the logs for other services than "ImportantApp"
                }
            }
        }
    }
}

Dropper

Dropper is a special sink, meaning that all logs are dropped without any exception. This is useful when you want to drop debugging logs.

filter {
    // ... parser

    sink {
        if (parsed.level == "DEBUG") {
            dropper {}
        } else {
            sampler {
                // ... configs
            }
        }
    }
}

Or if you have multiple filters, some of which are for extracting metrics, only one of them has to be persisted.

filter { // filter A: this is for persistence
    // ... parser

    sink {
        sampler {
            // .. sampler configs
        }
    }
}
filter { // filter B:
    // ... extractors to generate many metrics
    extractors {
        metrics {
            // ... metrics
        }
    }
    sink {
        dropper {} // drop all logs because they have been saved in "filter A" above.
    }
}

Enforcer

Enforcer is another special sink that forcibly samples the log. A typical use case of enforcer is when you have configured a sampler and want to save some logs forcibly, such as to save error logs even if the sampling mechanism has been configured.

filter {
    // ... parser

    sink {
        sampler {
            // ... sampler configs
        }
        if (parserd.level == "ERROR" || parsed.userId == "TestingUserId") { // sample error logs or testing users' logs (userId == "TestingUserId") even if the sampling strategy is configured
            enforcer {
            }
        }
    }
}