Skip to content
Anomalies detection system for periodic metrics.
Branch: master
Clone or download
zhaojkun Merge pull request #666 from zhaojkun/ignore_load_data
Ignore load data when rule is not trending related
Latest commit 121ecfb Jan 17, 2018
Type Name Latest commit message Commit time
Failed to load latest commit information.
Godeps Add deps Jul 1, 2017
algorithm set alert count even project silent or users' count is zero Sep 19, 2016
detector Change metric validation error log to debug level Jan 17, 2018
docker Add docker support Jan 16, 2017
filter Rule not trending related will not load history metrics Jan 17, 2018
static Fix user to project frontend bug Jul 21, 2017
statsd Add `Support Metric Types` to statsd README Feb 19, 2016
storage Fix for lint Sep 25, 2017
util add threshold for score Sep 19, 2016
vendor Add deps Jul 1, 2017
version Bump v0.2.4 Jun 2, 2016
Makefile Add rule import api May 14, 2017
main.go Add deps Jul 1, 2017


Banshee is a real-time anomalies(outliers) detection system for periodic metrics.

Build Status GoDoc Join the chat at



For example, a website api's response time is reported to banshee from statsd every 10 seconds:

20, 21, 21, 22, 23, 19, 18, 21, 22, 20, ..., 300

The latest 300 will be catched.


  • Designed for periodic metrics.
  • Dynamic threshold analyzation via 3-sigma.
  • Also supports fixed-threshold alert option.
  • Provides an alert rule management panel.
  • No extra storage services required.


  1. Go >= 1.5.
  2. Node and gulp.
  3. Statsd.

It is strongly recommended to use statsd as banshee client.


  1. Clone this repo and checkout to the latest release.
  2. Build binary via make.
  3. Build static files via make static.


$ ./banshee -c <config-filename>

Example configuration file is config/exampleConfig.yaml.

Statsd Integration

  1. Install statsd-banshee to forward metrics to banshee.

    $ cd path/to/statsd
    $ npm install statsd-banshee
  2. Add statsd-banshee to statsd backends in config.js:

     , backends: ['statsd-banshee']
     , bansheeHost: 'localhost'
     , bansheePort: 2015

Supported Metrics

  • timers: timer.mean_90.*, timer.upper_90.*, timer.count_ps.*.
  • counters: counter.*.
  • gauge: gauge.*.

Detection should work for any metric delimited by dots, but above types are better supported and are also recommended to use as banshee input.

Statsd-banshee would format banshee metric names before data sent out.

Web Panel Manual

Welcome to checkout the web panel manuals: English, 简体中文.


Banshee is a single-host program, its detection is fast enough in our case, we don't have a plan to expand it now.

We are using a Python script ( via fabric) to deploy it to remote host:

python -u hit9 -H remote-host:22 --remote-path "/service/banshee"


Just pull the latest tag release. Please don't use master branch directly, checkout to a tag instead.

Generally we won't release not-backward-compatiable versions, if any, related notes would be added to the changelog.

Alert Command

Banshee requires a command, normally a script to send alert messages.

It should be called from command line like this:

$ ./alert-command <JSON-String>

The JSON string example can be found at alerter/exampleCommand/echo.go.


But how do you really analyze the anomalous metrics? Via 3-sigma:

>>> import numpy as np
>>> x = np.array([40, 52, 63, 44, 54, 43, 67, 54, 49, 45, 48, 54, 57, 43, 58])
>>> mean = np.mean(x)
>>> std = np.std(x)
>>> (80 - mean) / (3 * std)
1.2608052883472445 # anomaly, too big
>>> (20 - mean) / (3 * std)
-1.3842407711224991 # anomaly, too small

For further implementation introduction, please checkout docs/

Network Protocol

If you are using statsd as banshee client, please checkout statsd-banshee.

The network protocol is line based:


Where the NAME should be a string, STAMP should be a timestamp integer in seconds, and the VALUE should be a float number.


Please checkout docs/

Docker Image

Please checkout docker/


Thanks to our contributors.


MIT Copyright (c) 2015 - 2016 Eleme, Inc.

You can’t perform that action at this time.