The Golang implementation for downsampling time series data algorithm
While monitoring the online system, there could be so many metrics' time series data will be stored into the Elasticsearch or NoSQL database for analysis. When the time passed, storing every piece of the historical data is not very effective way, and those huge data could impact the analysis performance and the cost of storage.
One of solution just simply delete the aged historical data(e.g. only keep the latest 6 months data), but there is a solution we can compressing those data to small size with good resolution.
Here is a demo shows how to downsamping the time series data from 7500 points to 500 points (Actually, you can downsample it to 200 or 300 points).
-
All of the algorithms are based on Sveinn Steinarsson's 2013 paper Downsampling Time Series for Visual Representation
-
This implementation refers to Ján Jakub Naništa's implementation by Typescript
-
The test data I borrow from one of python implementation which is here
Sveinn Steinarsson's paper mentioned 3 types of algorithm:
- Largest triangle three buckets (LTTB)
- Largest triangle one bucket (LTOB)
- Largest triangle dynamic (LTD)
You can find all of these implementation under src/downsampling directory.
Following the below instruction to compile and run this repo.
make
./build/bin/main
If everything goes fine, you will see the following message
2019/09/07 18:34:42 Reading the testing data...
2019/09/07 18:34:42 Downsampling the data from 7501 to 500...
2019/09/07 18:34:42 Downsampling data - LTOB algorithm done!
2019/09/07 18:34:42 Downsampling data - LTTB algorithm done!
2019/09/07 18:34:42 Downsampling data - LTD algorithm done!
2019/09/07 18:34:42 Creating the diagram file...
2019/09/07 18:34:43 Successfully created the diagram - ..../build/data/downsampling.chart.png
You can go to the ./build/data/ directory to check the diagram and the cvs files.
The diagram picture as below
- The first black chart at the top is the raw data with 7500 points
- The second, third, and fourth respectively are LTOB, LTTB, and LTD downsampling data with 500 points
- The last one at the bottom just put all together.
You can use the following makefile target to analyze the performance of these algorithms.
make prof
make bench
- The Billion Data Point Challenge by Uber Engineering team
- Visualize Big Data on Mobile by dduraz
- Sampling large datasets in d3fc by William Ferguson
- Downsampling algorithms by Adrian S. Tam
Enjoy it!