Self-scaling

Introduction

This service will be able to horizontally scale (up or down) the resources devoted to a specific service (inside a node) in a dynamic fashion, based on time series inference and custom logic.

Features

Self-scaling service shall store time series with the usage metrics of the serviceComponent of each active service in the host cluster. Deep learning techniques based on time series models are used to predict usage metrics and horizontally scale the resources dedicated to each service component. The software will be self-contained and will act accordingly to the dynamic behaviour of each service.

Place in architecture

When the administrator user enables the self-scaling controller service it automatically starts working. It accesses the metrics and stores them in its internal database, performs the deep learning process and infers to create the horizontal objects pod autoscalers dynamically. All this with pre-set values in the initial configuration.

../../_images/self-scaling_architecture.png

  • API REST: Contains the logic necessary to make GET and POST calls to intervene with the system behaviour, change default values or collect information.

  • Pod Self-Scalings Controller: Performs the collection of metrics and is responsible for storing the values in 15-minute intervals.

  • Database storage: Contains the history and predicted data of all serviceComponents of each active service. Also, contains the relative information about if each service are activated to infer or not.

  • Training module: Collects the raw data from the history databases and converts it to the format needed for the deep learning process. Executes the data predictions and stores them in a new database.

  • Inference module: Adds logic to the data in the future database and generates the inference process. Creates or replaces the horizontal pod autoscaler objects. Changes the previous values to the new ones based on the results obtained.

User guide

This component has a management API that provides a flask-based REST interface that can be interacted with to configure certain values. The url must include not only the address of the component, but also the action to be performed and the message body if necessary. The response shall include the requested information or the result of the execution of a command.

Method

Endpoint

Description

Payload (if need)

Response format

GET

/v1/services

Return services and serviceComponents

{“services”: [{“name”: “self-scaling”,”managed”: true,”serviceComponents”: [{“name”: “api”,”managed”: true}]}]}

POST

/v1/services

Update managed services

{“services”: [{“name”: “self-scaling”,”managed”: true,”serviceComponents”: [{“name”: “api”,”managed”: true}]}]}

[“services managed updates sucessfully”,”Content-Type not supported!”,”Invalid JSON”,”Managed services/serviceComponents must be true”]

GET

/v1/train-values

Return train values

{“Future_data”: “1”, “History_data”: “5”}

POST

/v1/train-values

Update train values

{“Future_data”: “1”, “History_data”: “5”}

[“Train values changed”,”Content-Type not supported!”,”Error in json body”,”Values must be positive numbers”]

GET

/v1/train

Execute the train

[“Train module executed successfully”,”Not components to train”,”Insufficient data”]

GET

/v1/inference

Execute the inference

[“Infence complete sucessfully”,”Error in execution Inference Module”]

GET

/version

Return version

{“component”: “self-scaling”,”version”: “1.0.0”}

GET

/v1/health

Return health status

{“status”: “healthy”}

GET

/v1/api-export

Return OpenAPI JSON format

{“openapi”: “3.0.0”,”info”: {…}}

Prerequisities

  1. Kubernetes cluster (e.g. microk8s, k8s).

  2. Helm package manager for kubernetes.

  3. Enable metrics server to get the metrics from cluster.

Installation

This component is provided as a Helm chart. Refer to specific deployment instructions.

helm install self-scaling <folder-to-self-scaling> --debug

Developer guide

This code is expected to be executed within a Helm chart, in a Kubernetes-governed platform. It has been also tested with Docker compose and directly over Ubuntu x64 distributions, with and without GPU NVIDIA processors. In case that developers aims at using the code directly over a given Operating System, non-virtualized, the code has been tested only in Ubuntu 20.04 machines, and hence we do not grant that it will work in any other OS.

This code is open source and can be freely used by the innovation and research community. In case that commits are to be made, the mantainer team (UPV) holds the rights to accept or deny them. Best practices are encouraged in the latter case.

Authors

Universitat Politècnica de València

License

This software is licensed under the Apache 2.0 license.

Notice (dependencies)

aerOS - Autonomous, scalablE, tRustworthy, intelligent European meta Operating System for the IoT edge-cloud continuum

Horizon Europe CL4-2021-DATA-01-05

Self-Scaling

Copyright 2022-2025 Universitat Politècnica de València

  1. Included Software

  2. Used Software

  1. List of licenses