This repository describes my implementation of the coding challenge after a first interview at Door2door Berlin. The project should be cloned from github here.

The project contains utilities, visualiser, simulator, webapp and test modules.

The simulator module is a clone of the simulator directory in the github page, with some changes as highlighted in the Edits in simulator paragraph. The visualiser module is used to generate locally-stored visualisations, and the webapp module shows the results in a web interface made in Flask. A utilities module is used to read the static data files. A test module contains some tests defined for the project.

The data directory contains two static data files. Addionally, this folder contains a contextily_cache directory to reduce the amount of requests to a tiling server.

• Simulator.path_to_stops is moved to be an instance variable.
• Added simulator/__init__.py to allow loading of the simulator as a module.
• simulator/requirements.txt are moved to requirements.txt and updated to match pandas==1.0.1 due to import errors on version 1.1.2
• Return type of simulator is not jsonified, e.g. get_random_points returns a geodataframe.
• Sets a default coordinate system in the generated geodataframe: EPSG:4326 (WGS84). This is necessary for getting map tiles from contextily, which operates in the current version in EPSG:3857 - Web mercator.
• If no points are found or points < n in the get_random_points() method, return the maximum amount of points.

Setting up the python environment is done as follows:

# make sure you have the right tools (assumes you have a working install of python)
pip install virtualenv-tools
# Initialize an empty venv that will exist in the current working directory as `venv`
python3 -m venv env
# enable the virtual environment
source env/bin/activate
# install requirements
pip install -r requirements.txt

The webapp module contains the web interface. All of the webpages are generated using the python flask module. I did not use React as I dont have any experience with this framework. The flask library is not necessarily good for production use, but can be used to interface with python applications easily.

Two directories are created in the webapp module:

• a templates directory to store the visualisation templates used to generate HTML. In this directory I have created four files: layout.html containing the essential layout (bootstrap) + a code block which we extend in the other pages. home.html is the landing page for the application. It contains a form that can be used to trigger a simulation & visualisation. visualisation.html is the visualisation page that loads the three visualisations and shows a button to perform a new simulation. A helper jinja file is given in _macros.jinja to facilitate generation of form elements.
• a static directory to store the generated images / maps.

Some other files are present:

• an __init__.py file to instantiate the module. This file creates the python flask application according to the application factory standard. This is mainly used to easily extend a webapp with new functionality.
• a config.py file containing the config of the web app. These are mostly paths and other configuration requirements. We use a csrf token to allow for form submission in the html page.
• a forms.py file containing the web form that is shown as a homepage when running the app. This form data can be altered by the user. A test is made to assert that the fields are of correct data type and filled.
• a routes.py file containing the required routing points for the webapp. In essence only two routes are necessary: a trigger_page route to initialize the simulation and generate the resulting images, and a visualise route to visualise the newly made results.

The webapp can be run from the project root using the previously defined environment.

source env/bin/activate
python app.py

A webserver will start on localhost:5000.

A docker image can be pulled and run directly from docker hub:

docker run --rm -p 5000:5000 --name mi-code-challenge fremmen/mi-code-challenge:latest

This docker image can also be locally built (architecture = amd64) from the project root:

docker build -t mi-code-challenge .

The locally built image can then be run with

docker run --rm -p 5000:5000 --name mi-code-challenge mi-code-challenge

Port 5000 is forwarded from container to host. A webserver will start on localhost:5000.

The docker image is based on python:3.8 and is 1.33GB in size. This is too large and a focus can be put on reducing the size. However, this is not the focus of the project and therefore it was ignored.

The modules can all be used as python3 modules. An import is done using e.g.

from visualiser.visualiser import Visualiser

A jupyter notebook is given API-showcase-notebook.ipynb that walks through the each of the modules in the application seperately. This file also contains documentation on each of the functions / classes.

To open this file the user should install jupyterlab:

pip install jupyterlab

Afterwards, the file can be shown using

jupyter notebook API-showcase-notebook.ipynb

The objective states that any of the simulations are within the boundaries of Berlin. A sanity check is made in the visualiser to visually confirm that the bbox is within bounds. Addionally, this sanity check should be provoked in the Simulator class to mathematically verify the solutions integrity. However modifying the Simulator class was out of scope.

A form is generated in the web interface at route home forces the user to input

• float in the x1,y1,x2,y2 fields
• int in the number_of_requests field.

If this is not true, the form will yield an error in the corresponding line and the request will not be forwarded until the user uses the correct datatype.

If a user is not using the webserver but instead the API, some checks are performed:

• In StaticDataReader: does the berlin_stops / berlin_bounds file exist? If it does not, return an empty geodataframe.
• in Visualiser.generate_overview_figure(): if the geodataframe is empty, ignore it.
• If the amount of data points from the Simulator is less than the n value, print "No Data found for XXX" on the close up image. The overview image still shows the total data + bounds.

Several tests have been defined in the tests module. More specifically for the BoundingBox, Simulator, StaticDataReader and Visualiser classes. Some of the test modules suppress Future and Deprication Warnings. The tests can be run as follows:

python -m unittest tests/xxxx.py

python -m unittest discover tests

• I replaced the React app with a Flask app. Flask is not good for production environments.
• As can be observed in webapp/routes.py, HTTPS is forced before each request. The flask app can be extended using such functionalities.
• Fetching the contextily web tiles can be very slow. An initial effort is made to reduce the number of required requests to the web tile server by hardcoding a cache path in the data/contextily_cache directory. Other options are available.

• The static data is re-read on every simulation. Both simulator and visualiser class re-read the data. This is ofcourse not necessary.

• The visualisations are generated to a directory, and this directory is cleaned up before a new request is completed when the webapp is used. This is ok, but it would be better to generate it on-the-fly. An option is to show a default Berlin visualisation and data can be highlighted by for example selection using mouse input.
• Two out of three visualisations use the commonly used matplotlib library. I used this library because of its heavy use in python development.
• The gmplot library is used for a last visualisation. There are many libraries such as this one. I just chose this one because of the dynamic map generation, it looks cool.
• A good extension to the visualisation module would be to use the output of the simulator.get_booking_distance_bins() method. However, this method is not related to the actual sampled points so it was ignored for this project.
• More data would mean better visualisations.
• Evaluating KPI is heavily dependent on the simulation data and defining a good metric should be done in a team discussion with extensive research, not by a single person. A good metric could be based on deriving a possible path planning, e.g. a solution to the commonly known travelling salesman problem that minimizes the travelled distance. However, perceived customer value - for example time spent in the vehicle - is also very important because the fuel consumption only relates to the monetary aspect of the service.
• No path is shown between sets of points. This is related to dropoff-pickup relationships.
• The Google Maps web page displays an error "This page can't load Google Maps correctly." due to no API key being present for this project.

• The Simulator requires a GeoPandas version of 0.5.0. In this version a FutureWarning is given for the initialisation of a GeoDataFrame with coordinate systems. This is according to this stackexchange post fixed in version > 0.7.0. However, touching the requirements for the simulator is out of scope. Fixing the warnings will be done after the migration to > 0.7.0 is done.
• A large overhead is added by the use of geodataframes, which in small datasets such as this one is not necessary.
• The Visualiser class is quite slow due to queries to the web tile servers. This can be improved using cached tiles.
• All of the classes are built to be easily extensible / maintainable, but ofcourse I am open to modifications to better match the other parts of the application.
• The Dockerfile is not optimized at all. Furthermore, the docker runs completely isolated and maybe a volume needs to be shared for the simulation results.

All of the cells in the API-showcase-notebook are timed using the %%timeit and %%prun commands and the following conclusions can be drawn:

Both matplotlib / contextily generated images are very slow, with a factor 500 slower when compared to the gmplot implementation.

This is likely due to the queries to the map tile servers which is inherently slow. When starting from no cached tiles, a single image generation can take up to 5.7 seconds. A simple solution was to add a caching directory.

After closer inspection it was observed that in the generation of the overview image around 750-850ms (36%) of execution time was due to image operations (opening, scaling, resampling, ...). Furthermore, the reprojection to a different coordinate system was also very slow ~340ms (15,5%). These operations can yield very large performance improvements if tackled properly.