!Newer rewritten version is available here: https://github.com/AICoE/prometheus-anomaly-detector
This python application has been written to deploy a training pipeline on OpenShift. This pipeline will at regular specified intervals collect new data directly from a prometheus instance and train a model on it regularly. This application also hosts a web page which can be used as a target for prometheus. This target currently serves 6 different metrics using two different prediction models (Prophet and Fourier Extrapolation).
To run this application you will need to install several libraries listed in the requirements.txt.
To install all the dependencies at once, run the following command when inside the directory:
pip install -r requirements.txt
After all the prerequisites have been installed, open the Makefile and you will see a list of required and optional variables in the beginning. The required variables will be used to communicate with the Prometheus and Storage end-points.
Populating the Makefile is the most important step, as you can use this to run the application on OpenShift, Docker or your local machine.
After setting up the credentials in your Makefile, run the following command to run a flask server which will regularly train and serve the predicted metrics as a prometheus target:
make run_model
After populating all the required variables, set the name for your docker app by changing the docker_app_name variable. Then run the following command to build the docker image.
make docker_build
This command uses the Dockerfile included in the repository to build an image. So you can use it to customize how the image is built.
After the image is successfully built, you can run the following command to run a flask server in a docker container, this command also specifies on which the predicted metrics are served which can be easily changed in the Makefile.
make docker_run
- In the Makefile set up the required variables, and then run the following command:
make oc_deploy
This will create a deployment on OpenShift and which after training the prophet model, will serve the predicted metrics as a web page (using the flask web server), these predicted metrics can later be easily collected by a prometheus instance.
Following is a sample web page view of what the metrics will look like:
# HELP process_virtual_memory_bytes Virtual memory size in bytes.
# TYPE process_virtual_memory_bytes gauge
process_virtual_memory_bytes 13.0
# HELP process_resident_memory_bytes Resident memory size in bytes.
# TYPE process_resident_memory_bytes gauge
process_resident_memory_bytes 31.0
# HELP process_start_time_seconds Start time of the process since unix epoch in seconds.
# TYPE process_start_time_seconds gauge
process_start_time_seconds 15.25
# HELP process_cpu_seconds_total Total user and system CPU time spent in seconds.
# TYPE process_cpu_seconds_total counter
process_cpu_seconds_total 69.88
# HELP process_open_fds Number of open file descriptors.
# TYPE process_open_fds gauge
process_open_fds 60.0
# HELP process_max_fds Maximum number of open file descriptors.
# TYPE process_max_fds gauge
process_max_fds 14.0
# HELP python_info Python platform information
# TYPE python_info gauge
python_info{implementation="CPython",major="3",minor="6",patchlevel="5",version="3.6.5"} 1.0
# HELP predicted_values_prophet Forecasted value from Prophet model
# TYPE predicted_values_prophet gauge
predicted_values_prophet{beta_kubernetes_io_arch="amd64",beta_kubernetes_io_os="linux",instance="cpt-0001.redhat.com",job="kubernetes-nodes",kubernetes_io_hostname="cpt-0001.redhat.com",node_role_kubernetes_io_compute="true",operation_type="create_container",provider="rhos",quantile="0.5",region="compute",size="small"} 32.99
# HELP predicted_values_prophet_yhat_upper Forecasted value upper bound from Prophet model
# TYPE predicted_values_prophet_yhat_upper gauge
predicted_values_prophet_yhat_upper{beta_kubernetes_io_arch="amd64",beta_kubernetes_io_os="linux",instance="cpt.redhat.com",job="kubernetes-nodes",kubernetes_io_hostname="cpt-0001.redhat.com",node_role_kubernetes_io_compute="true",operation_type="create_container",provider="rhos",quantile="0.5",region="compute",size="small"} 36.728885
# HELP predicted_values_prophet_yhat_lower Forecasted value lower bound from Prophet model
# TYPE predicted_values_prophet_yhat_lower gauge
predicted_values_prophet_yhat_lower{beta_kubernetes_io_arch="amd64",beta_kubernetes_io_os="linux",instance="cpt-0001.redhat.com",job="kubernetes-nodes",kubernetes_io_hostname="cpt-0001.redhat.com",node_role_kubernetes_io_compute="true",operation_type="create_container",provider="rhos",quantile="0.5",region="compute",size="small"} 27881.58691175386
# HELP predicted_values_fourier Forecasted value from Fourier Transform model
# TYPE predicted_values_fourier gauge
predicted_values_fourier{beta_kubernetes_io_arch="amd64",beta_kubernetes_io_os="linux",instance="cpt-0001.redhat.com",job="kubernetes-nodes",kubernetes_io_hostname="cpt-0001.redhat.com",node_role_kubernetes_io_compute="true",operation_type="create_container",provider="rhos",quantile="0.5",region="compute",size="small"} 29838.64724605837
# HELP predicted_values_fourier_yhat_upper Forecasted value upper bound from Fourier Transform model
# TYPE predicted_values_fourier_yhat_upper gauge
predicted_values_fourier_yhat_upper{beta_kubernetes_io_arch="amd64",beta_kubernetes_io_os="linux",instance="cpt-0001.redhat.com",job="kubernetes-nodes",kubernetes_io_hostname="cpt-0001.redhat.com",node_role_kubernetes_io_compute="true",operation_type="create_container",provider="rhos",quantile="0.5",region="compute",size="small"} 37111.31044977396
# HELP predicted_values_fourier_yhat_lower Forecasted value lower bound from Fourier Transform model
# TYPE predicted_values_fourier_yhat_lower gauge
predicted_values_fourier_yhat_lower{beta_kubernetes_io_arch="amd64",beta_kubernetes_io_os="linux",instance="cpt-0001.redhat.com",job="kubernetes-nodes",kubernetes_io_hostname="cpt-0001.redhat.com",node_role_kubernetes_io_compute="true",operation_type="create_container",provider="rhos",quantile="0.5",region="compute",size="small"} 29739.05799347848
- fbprohphet - Facebook's timeseries forecasting library
- requests - HTTP Library for python
- boto3 - AWS sdk for python
- pandas - High Performance Data Structure
- flask - A lightweight web application framework
- apscheduler - Python Scheduling library
- prometheus_client - Official Python client for Prometheus
- sortedcontainers - Pure python sorted simple data structures
- Anomaly Detection - Anomaly Detection Function by Natasha Frumkin
- Fourier Extrapolation Model - Fourier Extrapolation Model by Natasha Frumkin
- Serving Prometheus Metrics - Flask Server to host Prometheus metrics by Hema Veeradhi