wangchen615 / autoscaler_predictor_model

The data synthesizer, forecaster and predictor model server used together with KEDA scaler or cluster autoscaler to achieve cluster autoscaling with diurnal pattern workload

autoscaler_predictor_model

The data synthesizer, forecaster and predictor model server used together with KEDA scaler or cluster autoscaler to achieve cluster autoscaling with diurnal pattern workload

How to use the data synthesizer

Set up python virtual environment

python3 -m venv autoscaler_predictor_model/venv
source autoscaler_predictor_model/venv/bin/activate
pip install -r requirements.txt

Usage Instructions:

1. Configure the data synthesizer:

• Edit the config.ini file to set the peak and valley CPU/memory usage or requests/s load. You can also configure the number of weeks to generate the data, the peak hour in a day and the variations of usage/load across the same time in a day.

2. Generate Data:

python data_synthesizer.py

3. The data is generated in the data folder. The data is generated in CSV format with the following columns:

• timestamp: The timestamp in milliseconds since epoch
• cpu: The CPU usage in millicores
• memory: The memory usage in bytes
• requests: The number of requests per second The csv file name is in the format: <type>-<year>-<month>-<day>-<hour>-<minute>.csv and can be fed into the model server to fit the model. The json file name is in the format: <type>-<year>-<month>-<day>-<hour>-<minute>.json and can be fed into the model server to fit the model or to forecast the next datapoint.

How to use the model server

Build and Run the Docker image

podman build -t autoscaler-prediction-model-server .
podman run -p 5001:5001 autoscaler-prediction-model-server

Usage Instructions:

1. Fit the Model:

• Send a POST request to /fit-model with a CSV/JSON file containing timestamp and cpu/memory or requests columns.
• The server fits a polynomial model to this data.

An example curl command to fit the model:

curl -X POST http://localhost:5001/fit-model -F 'file=@/path/to/data.csv'

2. Get Predictions:

• Send a GET request to /predict with a type parameter (requests or resource) and a timestamp query parameter.
• The server returns the predicted cpu/memory or requests value(s) based on the fitted model.

An example curl command to get predictions:

curl -X GET "http://localhost:5001/predict?type=resource&timestamp=2023-03-15T14:30:00"
curl -X GET "http://localhost:5001/predict?type=requests&timestamp=2023-03-15T14:30:00"

3. Get Forecasts:

• Send a POST request to /forecast with a csv file or json object that contains any length of time series data.
• The server returns the predicted cpu/memory or requests value(s) based on SARIMA algorithm running on the given data. Example curl command to get forecasts from a csv file:

 curl -X POST http://127.0.0.1:5001/forecast \
-H "Content-Type: multipart/form-data" \
-F "file=@./data/requests-23-11-16-22-51.csv"

Example curl command to get forecasts from a json object:

curl -X POST http://127.0.0.1:5001/forecast \
-H "Content-Type: application/json" \
-d '[{"timestamp":1700092800000,"requests":279},{"timestamp":1700093700000,"requests":257},{"timestamp":1700094600000,"requests":230}]'

Notes:

• The fit-model endpoint expects a CSV file with two columns: timestamp and requests or three columes: timestamp, cpu and memory. Sample data can be found in the data folder.
• The predict endpoint requires a timestamp in a recognizable datetime format.
• The model fitting is simplistic and assumes a polynomial model. You may need to adjust the model fitting part based on your specific requirements.