Skip to main content

Extract anomalies from log files

Project description

Based on success logs, logreduce highlights useful text in failed logs. The goal is to save time in finding a failure’s root cause.

On average, learning run at 2000 lines per second, and testing run at 1300 lines per seconds.

How it works

logreduce uses a model to learn successful logs and detect novelties in failed logs:

  • Random words are manually removed using regular expression
  • Then lines are converted to a matrix of token occurrences (using HashingVectorizer),
  • An unsupervised learner implements neighbor searches (using NearestNeighbors).


This method doesn’t work when debug content is only included in failed logs. To successfully detect anomalies, failed and success logs needs to be similar, otherwise the extra informations in failed logs will be considered anomalous.

For example this happens with testr where success logs only contains ‘SUCCESS’.


  • Fedora:
sudo dnf install -y python3-scikit-learn
git clone
pushd logreduce
python3 develop --user
  • openSUSE:
sudo zypper install python3-scikit-learn
git clone
pushd logreduce
python3 develop --user
  • Pip:
pip install --user logreduce


Logreduce needs a baseline for success log training, and a target for the log to reduce.

Logreduce prints anomalies on the console, the log files are not modified:

"%(distance)f | %(log_path)s:%(line_number)d: %(log_line)s"

Local file usage

  • Compare two files or directories without building a model:
$ logreduce diff testr-nodepool-01/output.good testr-nodepool-01/
0.232 | testr-nodepool-01/  File "voluptuous/", line 370, in validate_mapping
0.462 | testr-nodepool-01/    raise er.MultipleInvalid(errors)
0.650 | testr-nodepool-01/  voluptuous.error.MultipleInvalid: required key not provided @ data['providers'][2]['cloud']
  • Compare two files or directories:
$ logreduce dir preprod-logs/ /var/log/
  • Or build a model first and run it separately:
$ logreduce dir-train sosreport.clf old-sosreport/ good-sosreport/
$ logreduce dir-run sosreport.clf new-sosreport/

Zuul job usage

Logreduce can query Zuul build database to train a model.

  • Extract novelty from a job logs:
$ logreduce job

# Reduce comparaison to a single project (e.g. for tox jobs)
$ logreduce job --project openstack/nova

# Compare using many baselines
$ logreduce job --count 10

# Include job artifacts
$ logreduce job --include-path logs/ http:/
  • Or build a model first and run it separately:
$ logreduce job-train --job job_name job_name.clf
$ logreduce job-run job_name.clf

Journald usage

Logreduce can look for anomaly in journald, comparing the last day/week/month to the previous one:

  • Extract novelty from last day journal:
$ logreduce journal --range day
  • Build a model using journal of last month and look for novelty in last week:
$ logreduce journal-train --range month good-journal.clf
$ logreduce journal-run --range week good-journal.clf

Server component

A server component may be deployed to build an anomaly database and produce dataset. This initial implementation is focused on Zuul builds and it doesn’t support importing arbritary files yet. More details in this specification

Components list:

  • logreduce-server: the REST and Gearman server
  • logreduce-worker: job executor
  • logreduce-client: client cli
  • logreduce-webui: logreduce web interface


  • PUT /anomaly/new: receive user report request from os_loganalyze
  • PUT /anomaly: import an anomaly report (json file generated by standalone cli)
  • GET /anomaly/{anomaly_id}: return an anomaly details
  • POST /anomaly/{anomaly_id}: update anomaly status
  • POST /anomaly/{anomaly_id}/logfile/{logfile_id}: update scores
  • DELETE /anomaly/{anomaly_id}/logfile/{logfile_id}: remove a file
  • GET /anomalies: return the list of anomalies
  • GET /status: return the list of worker jobs


This package contains tests data for different type of log such as testr or syslog. Each tests includes a pre-computed list of the anomalies in log failures.

This package also includes a command line utility to run logreduce against all tests data and print a summary of its performance.

Test format

Each tests case is composed of:

  • A .good file (or directory) that holds the baseline
  • A .fail file (or directory)
  • A info.yaml file that describe expected output:
threshold: float # set the distance threshold for the test
  - optional: bool  # to define minor anomalies not considered false positive
    lines: |        # the expected lines to be highlighted


To run the evaluation, first install logreduce-tests:

git clone
pushd logreduce-tests
python3 develop --user

logreduce-tests expect tests directories as argument:

$ logreduce-tests tests/testr-zuul-[0-9]*
[testr-zuul-01]: 100.00% accuracy,  5.00% false-positive
[testr-zuul-02]:  80.00% accuracy,  0.00% false-positive
Summary:  90.00% accuracy,  2.50% false-positive

Add –debug to display false positive and missing chunks.


  • Add terminal colors output
  • Add progress bar
  • Better differentiate training debug from testing debug
  • Add a starting log line and report written
  • Add tarball traversal in utils.files_iterator
  • Add logstash filter module
  • Improve tokenization tests


  • Add daemon worker mode with MQTT event listener
  • Discard files that are 100% anomalous
  • Report mean diviation instead of absolute distances
  • Investigate second stage model


Contribution are most welcome, use git-review to propose a change. Setup your ssh keys after sign in

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Filename, size & hash SHA256 hash help File type Python version Upload date
logreduce-0.4.0-py2.py3-none-any.whl (63.9 kB) Copy SHA256 hash SHA256 Wheel py2.py3
logreduce-0.4.0.tar.gz (157.9 kB) Copy SHA256 hash SHA256 Source None

Supported by

Elastic Elastic Search Pingdom Pingdom Monitoring Google Google BigQuery Sentry Sentry Error logging AWS AWS Cloud computing DataDog DataDog Monitoring Fastly Fastly CDN SignalFx SignalFx Supporter DigiCert DigiCert EV certificate StatusPage StatusPage Status page