Welcome to Skyline.
What is Skyline?¶
Skyline is your ears and eyes, it is remarkably good at telling you when a state changes. It does not just chew bubblegum, it blows bubbles too.
But really Skyline is … For those interested in anomaly detection and deflection in streamed time series data.
Anomaly deflection. The obvious next evolution in the use of all the anomaly detection data?
Skyline is a Python based anomaly detection/deflection stack that analyses, anomaly detects, deflects, fingerprints and learns vast amounts of streamed time series data. Skyline has a number of isolated modules/apps that:
- ingests streamed metric time series data - skyline/horizon
- use a
`CONSENSUS`of 3-sigma algorithms to detect anomalies on batch processed, streamed metric time series data - skyline/analyzer - anomaly detector
- Handle large and small seasonality in the data - skyline/mirage - anomaly deflector and detector
- You can train it on what is NOT anomalous and it learns - skyline/ionosphere - anomaly deflector
- It records all your anomalies - skyline/panorama - anomaly memory
- It shows you all your data - skyline/webapp - anomaly view
Seeing as we desire our metrics to be not anomalous most of the time and we want to know when they ARE anomalous and given the fact that we try and build systems that try to behave within not anomalous bounds so they perform well, due to this we have:
- A lot of metric time series data that are not anomalous most of the time.
- A lot of data to train a system on what is NOT anomalous given a time series data set, rather than simply focusing on what is anomalous, also focusing on what is not anomalous.
To achieve this Skyline implements a novel time series similarities comparison algorithm and a boundary layers methodology that generates fingerprints of time series data using the sum of the values of features of the time series which have been extracted using the tsfresh features extraction package - https://github.com/blue-yonder/tsfresh and evaluation against boundary layer algorithms to determine whether a 3-sigma triggered anomaly is actually a normal, known pattern in the data.
The Skyline-Ionosphere-Tsfresh Time Series Similarities Comparison Algorithm - SITTSSCA first coined here :) compares the generated fingerprints of the two time series and can determine if they closely resemble each other in terms of:
- of the amount of “power/energy”, range and “movement” there is within the time series data set somewhat like RMS - Erol Kalkan from United States Geological Survey, “Another approach to compute the differences between two time series is moving window root-mean-square. RMS can be run for both series separately. This way, you can compare the similarities in energy (gain) level of time series. You may vary the window length for best resolution.” (https://www.researchgate.net/post/How_can_I_perform_time_series_data_similarity_measures_and_get_a_significance_level_p-value) http://stackoverflow.com/questions/5613244/root-mean-square-in-numpy-and-complications-of-matrix-and-arrays-of-numpy
The Skyline-Ionosphere-Tsfresh Time Series Similarities Comparison Algorithm compares how close the fingerprint values are as a percentage and varying this percentage variable will either focus the algorithm with greater precision, the closer to 0% the parameter gets, the perfect match (or possibly a mirror match too - unkonwn/untested) or it will incrementally increase the tolerance as the percentage variable increases and the matching will become less and less reliable.
However there is a sweet spot and here SITTSSCA works extremely well :)
Added to SITTSSCA is an optional layer of simple boundary algorithms that are user defined during the operator training interaction with Skyline, where the operator augments the SITTSSCA results with boundaries that describe the expected norm within the time series. Very similar to being able to describe the Active Brownian Motion of a time series - https://github.com/blue-yonder/tsfresh/pull/143#issuecomment-272314801
This results in an anomaly detection/deflection system which enables the user to very simply label time series and train Skyline on the peaks and troughs and the expected Active Brownian Motion or best effort thereof.
However it takes a little effort on your part to train Skyline, however with the effort, Skyline is very good at doing anomaly detection and deflection.
With your help. There is no easy anomaly detection or deflection, but there is some reward with a bit of effort.
What you need¶
You must be interested in doing real time anomaly detection on vast amounts of time series data streams on a reasonable Linux VM (or in a container/s if you were really committed).
You are still here.
You use Graphite as your time series data source? No?
Currently Skyline needs Graphite, if you do not already have Graphite please consider setting Graphite up and if you like that, come back.
If you are still here and reading this, then maybe you are serious about installing and trying Skyline. In which case a word of warning, continuing from this point forward, will require a LOT of hours of your time.
Skyline is not some Python data science anomaly detection library, it is a full blown production grade anomaly detection stack. Although certain aspects of Skyline may have interest to the data science community.
Too much effort?¶
Try our managed service. We offer a managed version of Skyline for people that do not have a vast amount of time to spare. You’ll get access to unreleased features and support from developers that have honed numerous Skyline integrations to alert on important metrics.
We are looking for test partners as the product is currently in beta phase. Send us an email at firstname.lastname@example.org
Places are filling up quickly!
A brief history¶
Skyline was originally open sourced by Etsy as a real-time anomaly detection system. It was originally built to enable passive monitoring of hundreds of thousands of metrics, without the need to configure a model/thresholds for each one, as you might do with Nagios. It was designed to be used wherever there are a large quantity of high-resolution time series which need constant monitoring. Once a metric stream was set up from Graphite, additional metrics are automatically added to Skyline for analysis, anomaly detection, alerting and briefly published in the Webapp frontend. github/etsy stopped actively maintaining Skyline in 2014.
Skyline - as a work in progress¶
Etsy found the “one size fits all approach” to anomaly detection wasn’t actually proving all that useful to them.
There is some truth in that in terms of the one size fits all methodology that Skyline was framed around. With hundreds of thousands of metrics this does make Skyline fairly hard to tame, in terms of how useful it is and tuning the noise is difficult. Tuning the noise to make it constantly useful and not just noisy, removes the “without the need to configure a model/thresholds” element somewhat.
It has been generally accepted now that a basic 3-sigma anomaly detection implementation is not generally useful in the operations and machine metrics space.
|David Gildeh:||“I still remember taking Skyline and applying it to one of our customer’s metrics, and turning 100,000 metrics into 10,000 anomalies. It just created more noise from the noise.” https://blog.outlyer.com/what-good-is-anomaly-detection|
This is still true of Skyline today, it will still detect the 10000 anomalies and it should.
So why continue developing Skyline?
To try and make it better and more useful. 3-sigma anomaly detection works, but it works too well. Therefore there is an opportunity to see if it is possible to augment 3-sigma methods with additional analyses with new and different techniques, including the use of historic data in real time, to be more useful and provide additional insight into related time series data. One of the key paradigm shifts that is perhaps needed is to change the mindset that anomaly detection and alerting are synonymous with each other or related in any way, which seems to be general public opinion. Skyline does anomaly detection, anomaly deflection, training and learning, and alerting is simply a byproduct of this analysis pipeline, if you want to enable it.
The first way to make Skyline MUCH better than the manner it was implemented and framed by Etsy, is to NOT try and use it to alert on 1000s of metrics in the first place. Using Skyline as a scapel for alerting and a sword for anomaly detecting, rather than using it as a sword for anomaly detecting AND alerting.
Within in this paradigm, Skyline is still essentially 3-sigma based, however now being augmented with additional analysis and methods, Skyline has been much improved in many ways and is very useful at doing anomaly detection, recording anomalies, correlating and alerting and training on your KEY metrics. The ongoing development of Skyline has been focused on improving Skyline in the following ways:
- Improving the anomaly detection methodologies used in the 3-sigma context.
- Extending Skyline’s 3-sigma methodology to enable the operator and Skyline to handle seasonality in metrics.
- The addition of an anomalies database for learning and root cause analysis.
- Adding the ability for the operator to teach Skyline and have Skyline learn things that are NOT anomalous using a time series similarities comparison method based on features extraction and comparison using the tsfresh package. With Ionosphere we are training Skyline on what is NOT anomalous, rather than focusing on what is anomalous. Ionosphere allows us to train Skyline as to what is normal, even if normal includes spikes and dips and seasonality. After all we have some expectation that most of our metrics would be NOT anomalous most of the time, rather than anomalous most of the time. So training Skyline what is NOT ANOMALOUS is more efficient than trying to label anomalies.
- Adding the ability to Skyline to determine what other metrics are related to an anomaly event using cross correlation analysis of all the metrics using Linkedin’s luminol library when an anomaly event is triggered and recording these in the database to assist in root cause analysis.
The architecture/pipeline works very well at doing what it does. It is solid and battle tested..
Skyline is FAST!!! Faster enough to handle 10s of 1000s of time series in near real time. In the world of Python, data analysis, R and most machine learning, Skyline is FAST. Processing and analysing 1000s and 1000s of constantly changing time series, every minute of every day and it can do it in multiple resolutions, on a fairly low end commodity server.
Is Skyline better than other things at anomaly detection? Unknown. The development of Skyline is not focused on making it be better than other things or the best, it is focused on trying to make Skyline better than it was and currently is. Unfortunately Skyline no longer fits the NAB benchmark method as it’s methods operate exclusively in the real time arena on real time data, historic data and trained patterns and this could not be bolted into a NAB test and would violate the NAB benchmark requirements.
The new look of Skyline apps¶
- Horizon - feed metrics to Redis via a pickle input from Graphite/s
- Analyzer - analyses metrics with 3-sigma algorithms
- Mirage - analyses specific metrics at a custom time range with 3-sigma algorithms
- Boundary - analyses specific time series for specific conditions
- Crucible - store anomalous time series resources and adhoc analysis of any time series
- Panorama - anomalies database, historical views and root cause analysis
- Webapp - frontend to view current and historical anomalies, training data, features profiles, layers, matches and can browse Redis with rebrow and you manage Skyline’s learning with it
- Ionosphere - time series fingerprinting and learning
- Luminosity - Cross correlation of metrics for root cause analysis
Skyline is still a near real time anomaly detection system, however it has various modes of operation that are modular and self contained, so that only the desired apps need to be enabled, although the stack is now much more useful with them all running. This modular nature also isolated the apps from one another in terms of operation, meaning an error or failure in one does not necessarily affect others.
Skyline can now be feed/query and analyse time series on an ad hoc basis, on the fly. This means Skyline can now be used to analyse and process static data too, it is no longer just a machine/app metric fed system, if anyone wanted to use it to analyse historic data.
See whats-new for a comprehensive overview and description of the latest version/s of Skyline.
It must be stated the original core of Skyline has not been altered in any way,
other than some fairly minimal Pythonic performance improvements, a bit of
optimization in terms of the logic used to reach
settings.CONSENSUS and a
package restructure. In terms of the original Skyline Analyzer, it does the
same things just a little differently, hopefully better and a bit more.
There is little point in trying to improve something as simple and elegant in methodology and design as Skyline, which has worked so outstandingly well to date. This is a testament to a number of things, in fact the sum of all it’s parts, Etsy, Abe and co. did a great job in the conceptual design, methodology and actual implementation of Skyline and they did it with very good building blocks from the scientific community.
The architecture in a nutshell¶
Skyline uses to following technologies and libraries at its core:
- Python - the main skyline application language - Python
- Redis - Redis an in-memory data structure store
- numpy - NumPy is the fundamental package for scientific computing with Python
- scipy - SciPy Library - Fundamental library for scientific computing
- pandas - pandas - Python Data Analysis Library
- mysql/mariadb - a database - MySQL or MariaDB
- rebrow - Skyline uses a modified port of Marian Steinbach’s excellent rebrow
- tsfresh - tsfresh - Automatic extraction of relevant features from time series
- memcached - memcached - memory object caching system
- pymemcache - pymemcache - A comprehensive, fast, pure-Python memcached client
- luminol - luminol - Anomaly Detection and Correlation library