Behavior-Driven Forecasting for Vertical Auto-Scaling in Container Cloud
Theta-Scan is a method to detect and forecast behavior patterns on CPU/Memory demand in containerized applications. It is based on the deseasonalized ThetaForecaster algorithm, a time-series method that only requiresthe observation window and the expected periodicity as parameters for a prediction. Theta-Scan is in charge of finding the cycle from the observed window, indicating stationarity/trending/periodic patterns, and obtaining a forecaster model without additional computational cost.
The ThetaScan package contains the necessary code to 1) generate synthetic traces with stationarity/trending/periodic patterns, 2) process the trace and compute resource limit recommendations and 3) visualization.
- ThetaScan.py: class containing the needed functions to detect behaviours, scan periods and compute ThetaScan recommendations
- ThetaModel.py : class implementing the Theta Model, the best-fitted deseasonalized and detrended Simplistic Exponential Smoothing (Holt-Winters)
- utilities.py : support functions to generate synthetic traces, process and visualize them.
example.py: an example with generation and recommendation using the ThetaScan method
Theta Scan Autoscaling.ipynb: a jupyter notebook with generation and forecast of resources with ThetaScan. Two approaches are included: using a 5-minute fixed window for provisioning resources and a dynamic window dependent on the detected period.
To use the package import it into your python script:
from ThetaScan.ThetaScan import *
The required packages are: numpy, scipy, random, matplotlib
ThetaScan(): constructor of the class.scan(trace): scans the trace to detect periods. Returns the detected period and the smape loss for each of the scanned periods.detect_adf_stationarity(trace): performs the Augmented Dickey-Fuller Stationarity test on the trace. Returns true if stationary.detect_stationarity(trace): performs our stationarity test based on the period errors. Returns true if stationary.detect_trend(trace, period): retrieves the trend of the trace (slope, detrended and fitted trend).detect_behaviour(trace): detects the type of behaviour that the trace follows (trending/periodic/stationary).next_step_forecasting_theta(trace, period, n_steps): produces a Theta forecast given the detected period and the steps to produce.forecast_segment(trace, window_size, i, observation_window, prev_usage): detects the behaviour of the trace and produces the next forecast and provisioning limit.dynamic_forecast_segment(trace, idx_init, idx_end, default_window, observation_window,prev_usage): same as previous but with dynamic window.recommend(trace, window_size, observation_window): produces the complete resource recommendation for a resource usage trace.dynamic_recommend(trace, observation_window): same as previous with dynamic time window.
ThetaModel(): constructor of the class.fit(y, sp): deseasonalization, detrending and SES fitting.forecast (n_forecast): forecast, resesonalization and retrending.