ugianka/capstone1

AAVAIL revenue predictor service

Presentations

The EDA and model selection presentations are in the presentations directory under the project root directory

Installation

install in docker

There is a dockerfile on the root of the project to create the docker image. Be sure to have docker running on your machine. Then execute from the root directory:

docker build -t <image-name> .

this will create an image with name <image-name> in your docker registry.

To run the image type:

docker run -d -p 5000:5000 --name <container-name> <image-name>

<container-name> is the name you want to give to the container.

The --name <aname> parameter is optional if you don't specify it docker will create a name for the container.

The -p 5000:5000 will create a NAT in your machine on port 5000 to the port 5000 of the container. You will be able to send requests to the flask server in the container addressing your machine like this for example:

curl -X GET  http://127.0.0.1:5000/train

to verify that the container is running:

docker ps

this should show something like:

CONTAINER ID        IMAGE                  COMMAND               CREATED             STATUS              PORTS                    NAMES
be97439e4684        aavail_predict:1.0.9   "./start-server.sh"   5 seconds ago       Up 3 seconds        0.0.0.0:5000->5000/tcp   aavail
(

Service interface

These are the path of the different services, all services accept GET method and receive all parameters in query string

path	message	parameters	description
/train	GET	train_params	train all the models in production mode
/predict	GET	predict_params	predict the revenue for the month after the date passed as parameter
/getLog	GET	getlog_params	get the logs for the date passed as parameter

where

train_params are:

param name	description
env	can be either test or prod if not present env will be set to prod

predict_params are:

param name	description
country	the name of the country is all lowercase and spaces must be replaced by _
year	the year of the date for wich we want the prediction of the revenue of the next month
month	the month of the date
day	the day of the date
env	can be either test or prod if not present env will be set to prod

getlog_params are:

param name	description
type	either pred for prediction logs or train for train logs
date	the date in ISO format like 2020-12-01 for December 1st 2020

service response

all services respond with a json. Status code of the http response is always 200 even on failure. The field result in the JSON response is 0 if the API has executed without errors -1 otherwise . In case of error also the field error is present with a description of the error

services invocation examples

example of train

this will train all models for the top 10 countries

curl -X GET http://127.0.0.1:5000/train?env=test

example of prediction

for example if we want the prediction of the revenue for november 2018 for the country United Kingdom:

curl -X GET "http://127.0.0.1:5000/predict?country=united_kingdom&year=2018&month=10&day=31&env=test"

example of log fetch

this will retrieve all the prediction logs for the day: 2018-01-01. The date is relative to the day in which operations (predict or train) have been done .

curl -X GET "http://127.0.0.1:5000/getLog?type=pred&date=2020-12-01&env=test"

Unit Tests

Unit Tests are runned by the unit_test.sh shell in the root directory. Once you have run the application in a docker container you can execute unit tests with the command:

docker exec -ti <container-name> /usr/src/aavail_predict/unit_test.sh

where <container-name> is the name or the id of the container running the application

performance monitoring

Performances monitoring is done through the script:

./start_momitoring.sh

that is placed on the root directory of the project. This execute the python file monitoring.py that creates in monitor directory :

monitor_statistics.csv

which is a csv that contains for each day the monitor has run the mean squared error of the prediction of the 30 mext days for each date in the train data with respect to the actual revenues. As the train data will be populated and the model retrained a comparison of this mse could be done.

It creates also a csv file for each day the monitor is run that contains the detail of the prediction and the actual 30 days revenues and their difference.

It ctreates also an image of the distribution of the differences between predictions and actual.