system-config/doc/source/logstash.rst
Andrea Frittoli (andreaf) 756471608d Add periodic pipeline test result to infra docs
We now capture test results for the the periodic pipeline as well,
so changing the docs to reflect that.

Change-Id: I21f86157a85346d15ae703e6fab7ff993494b712
2015-12-15 10:17:10 +00:00

12 KiB

title

Logstash

Logstash

Logstash is a high-performance indexing and search engine for logs.

At a Glance

Hosts
Puppet
Configuration
  • modules/openstack_project/files/logstash
  • modules/openstack_project/templates/logstash
Projects
Bugs

Overview

Logs from Jenkins test runs are sent to logstash where they are indexed and stored. Logstash facilitates reviewing logs from multiple sources in a single test run, searching for errors or particular events within a test run, as well as searching for log event trends across test runs.

System Architecture

There are four major layers in our Logstash setup.

  1. Log Pusher Scripts. Subscribes to the Jenkins ZeroMQ Event Publisher listening for build finished events. When a build finishes an event is received from Jenkins which is then converted into Gearman jobs specific to that event for each log file we care about. These jobs trigger Gearman workers that then fetch the logs generated by that build, chop them up, annotate them with Jenkins build info and finally sends them to a Logstash indexer process.
  2. Logstash Indexer. Reads these log events from the log pusher, filters them to remove unwanted lines, collapses multiline events together, and parses useful information out of the events before shipping them to ElasticSearch for storage and indexing.
  3. ElasticSearch. Provides log storage, indexing, and search.
  4. Kibana. A Logstash oriented web client for ElasticSearch. You can perform queries on your Logstash logs in ElasticSearch through Kibana using the Lucene query language.

Each layer scales horizontally. As the number of logs grows we can add more log pushers, more Logstash indexers, and more ElasticSearch nodes. Currently we have multiple Logstash worker nodes that pair a log pusher with a Logstash indexer. We did this as each Logstash process can only dedicate a single thread to filtering log events which turns into a bottleneck very quickly. This looks something like:

jenkins
   |
   |
gearman-client ---------------
/    |    \                 |
/     |     \                |
gearman gearman gearman    subunit gearman
worker1 worker2 worker3       worker01
|      |      |               |
logstash logstash logstash         |
indexer1 indexer2 indexer3         |
\      |      /          subunit2sql
\     |     /                DB
elasticsearch
cluster
   |
   |
kibana

Log Pusher

This is a pair of simple Python scripts. The first listens to Jenkins build events and converts them into Gearman jobs and the second performs Gearman jobs to push log files into logstash.

Log pushing looks like this:

  • Jenkins publishes build complete notifications.
  • Receive notification from Jenkins and convert to Gearman jobs.
  • Using info in the Gearman job log files are retrieved.
  • Log files are processed then shipped to Logstash.

Using Gearman allows us to scale the number of log pushers horizontally. It is as simple as adding another process that talks to the Gearman server.

If you are interested in technical details the source of these scripts can be found at

Subunit Gearman Worker

Using the same mechanism as the Log pushers there is an additional class of gearman worker that takes the subunit output from test runs and stores them in a subunit2SQL database. Right now this is only done with the subunit output from gate and periodic queue tempest runs.

If you are interested in technical details the source of this script can be found at:

Logstash

Logstash does the heavy lifting of squashing all of our log lines into events with a common format. It reads the JSON log events from the log pusher connected to it, deletes events we don't want, parses log lines to set the timestamp, message, and other fields for the event, then ships these processed events off to ElasticSearch where they are stored and made queryable.

At a high level Logstash takes:

{
  "fields" {
    "build_name": "gate-foo",
    "build_numer": "10",
    "event_message": "2013-05-31T17:31:39.113 DEBUG Something happened",
  },
}

And turns that into:

{
  "fields" {
    "build_name": "gate-foo",
    "build_numer": "10",
    "loglevel": "DEBUG"
  },
  "@message": "Something happened",
  "@timestamp": "2013-05-31T17:31:39.113Z",
}

It flattens each log line into something that looks very much like all of the other events regardless of the source log line format. This makes querying your logs for lines from a specific build that failed between two timestamps with specific message content very easy. You don't need to write complicated greps instead you query against a schema.

The config file that tells Logstash how to do this flattening can be found at modules/openstack_project/templates/logstash/indexer.conf.erb

This works via the tags that are associated with a given message.

The tags in modules/openstack_project/templates/logstash/indexer.conf.erb are used to tell logstash how to parse a given file's messages, based on the file's message format.

When adding a new file to be indexed to modules/openstack_project/files/logstash/jenkins-log-client.yaml at least one tag from the indexer.conf.erb file should be associated with the new file. One can expect to see '{%logmessage%}' instead of actual message data if indexing is not working properly.

In the event a new file's format is not covered, a patch for modules/openstack_project/templates/logstash/indexer.conf.erb should be submitted with an appropriate parsing pattern.

ElasticSearch

ElasticSearch is basically a REST API layer for Lucene. It provides the storage and search engine for Logstash. It scales horizontally and loves it when you give it more memory. Currently we run a multi-node cluster on large VMs to give ElasticSearch both memory and disk space. Per index (Logstash creates one index per day) we have N+1 replica redundancy to distribute disk utilization and provide high availability. Each replica is broken into multiple shards providing inceased indexing and search throughput as each shard is essentially a valid mini index.

To check on the cluster health, run this command on any es.* node:

curl -XGET 'http://localhost:9200/_cluster/health?pretty=true'

Kibana

Kibana is a ruby app sitting behind Apache that provides a nice web UI for querying Logstash events stored in ElasticSearch. Our install can be reached at http://logstash.openstack.org. See query-logstash for more info on using Kibana to perform queries.

subunit2SQL

subunit2SQL is a python project for taking subunit v2 streams and storing them in a SQL database. More information on the subunit protocol can be found here: https://github.com/testing-cabal/subunit/blob/master/README

subunit2sql provides a database schema, several utilities for interacting with the database, and a python library to build tooling on top of the database. More information about using subunit2sql can be found at: http://docs.openstack.org/developer/subunit2sql/

Our instance of the subunit2SQL database is running on a MySQL database and has been configured to be remotely accessible to allow for public querying. The public query access is provided with the following credentials:

username: query
password: query
hostname: logstash.openstack.org
database name: subunit2sql
database port: 3306

simpleproxy

Simpleproxy is a simple tcp proxy which allows forwarding tcp connections from one host to another. We use it to forward mysql traffic from a publicly accessible host to the trove instance running the subunit2sql MySQL DB. This allows for public access to the data on the database through the host logstash.openstack.org.

Querying Logstash

Hop on over to http://logstash.openstack.org and by default you get the last 15 minutes of everything Logstash knows about in chunks of 100. We run a lot of tests but it is possible no logs have come in over the last 15 minutes, change the dropdown in the top left from Last 15m to Last 60m to get a better window on the logs. At this point you should see a list of logs, if you click on a log event it will expand and show you all of the fields associated with that event and their values (note Chromium and Kibana seem to have trouble with this at times and some fields end up without values, use Firefox if this happens). You can search based on all of these fields and if you click the magnifying glass next to a field in the expanded event view it will add that field and value to your search. This is a good way of refining searches without a lot of typing.

The above is good info for poking around in the Logstash logs, but one of your changes has a failing test and you want to know why. We can jumpstart the refining process with a simple query.

@fields.build_change:"$FAILING_CHANGE" AND @fields.build_patchset:"$FAILING_PATCHSET" AND @fields.build_name:"$FAILING_BUILD_NAME" AND @fields.build_number:"$FAILING_BUILD_NUMBER"

This will show you all logs available from the patchset and build pair that failed. Chances are that this is still a significant number of logs and you will want to do more filtering. You can add more filters to the queriy using AND and OR and parentheses can be used to group sections of the query. Potential additions to the above query might be

  • AND @fields.filename:"logs/syslog.txt" to get syslog events.
  • AND @fields.filename:"logs/screen-n-api.txt" to get Nova API events.
  • AND @fields.loglevel:"ERROR" to get ERROR level events.
  • AND @message"error" to get events with error in their message. and so on.

General query tips:

  • Don't search All time. ElasticSearch is bad at trying to find all the things it ever knew about. Give it a window of time to look through. You can use the presets in the dropdown to select a window or use the foo to bar boxes above the frequency graph.
  • Only the @message field can have fuzzy searches performed on it. Other fields require specific information.
  • This system is growing fast and may not always keep up with the load. Be patient. If expected logs do not show up immediately after the Jenkins job completes wait a few minutes.

crm114

In an effort to assist with automated failure detection, the infra team has started leveraging crm114 to classify and analyze the messages stored by logstash.

The tool utilizes a statistical approach for classifying data, and is frequently used as an email spam detector. For logstash data, the idea is to flag those log entries that are not in passing runs and only in failing ones, which should be useful in pinpointing what caused the failures.

In the OpenStack logstash system, crm114 attaches an error_pr attribute to all indexed entries. Values from -1000.00 to -10.00 should be considered sufficient to get all potential errors as identified by the program. Used in a kibana query, it would be structured like this:

error_pr:["-1000.0" TO "-10.0"]

This is still an early effort and additional tuning and refinement should be expected. Should the crm114 settings need to be tuned or expanded, a patch may be submitted for this file, which controls the process: https://git.openstack.org/cgit/openstack-infra/puppet-log_processor/tree/files/classify-log.crm