DCSIL

We will use EC2 to run our service using docker containers when we need to achieve more scale. Running in EC2 will allow us to easily create multiple servers for the application, front them with an elastic load balancer and use auto scaling to scale them as necessary. It allows us to achieve scale while also staying within the AWS cloud, which cannot be done with any other service.

Using docker will allow us to set up the local dev environment easily. We will be able to spin up our service, including with a local db for testing with a single command. It also gives us an easy to manage isolated environment. Docker is the easiest container service to use because of its documentation and wide spread support.

The simplest way to manage our codebase among our team members. Git and Github are also the most familiar version control tools among our team members.

Mailgun is a service that will allow us to send emails with offers directly from our service. We chose this over trying to set up our own SMTP server because it is much faster and easier to use. Speed is important for us because as a startup we want to get to MVP as fast as possible. Additionally, Mailgun gives us 20k free emails, which will be enough for creating our initial MVP.

Amazon RDS will be used to host our production DB. It is Amazon's offering for hosting mySQL databases. It is simple to connect it to EC2 or Lightsail instances. It is advantageous for us to use because it offers 20GB of free storage which will be enough for our machine learning model data. There are many alternative options to the RDS, but this is the only that keeps all our infrastructure in the same place and the simplest to set up given the rest of our tech stack.

Amazon S3 will be used for storing the ML models that we create. Amazon S3 is extremely simple to use and very cheap for the limited scale we will be using it for. Choosing Amazon S3 also allows us to stay within the AWS cloud and keeping everything together is convenient when developing.

Redis will be used as a caching layer between the database and our server. The requests that we plan to cache are very simple single entry lookups. Redis works extremely well as a cache because it provides sub millisecond response times and is extremely simple to use. We want to use a cache because it should reduce the requests to the database, which should reduce both cost of operation and latency. Another popular alternative to Redis is Memcached, which can provide similar benefits. We chose Redis because Amazon provides a fully managed service for Redis which can reduce the work needed to manage and set up the Redis cache and also provides replication that is unavailable with managed Memcached service. This will provide an edge as we scale our infrastructure and focus on reliably delivering fast performance.

We will be getting data in the form of CSVs. Because the data in a CSV is highly structured, it will be easy to create schemas and it works well in a SQL database as opposed to noSQL. For a SQL database, both mySQL and Postgres are very viable options. Both of them are highly performant, definitely enough for our application, even if we needed to scale drastically. Postgres does include some extra features over mySQL such as table inheritance and function overloading. However, the extra features are not advantageous to us given our database use case. Because both databases seemed to suit our use case perfectly, we chose to use mySQL simply because it is more familiar tech within our team.

NodeJS will be used as the runtime for our Javascript backend. There are some alternatives for JavaScript runtimes, such as Deno. However, NodeJS is not only the most popular by far, but also the most familiar to our team. It is incredibly simple to install and set up. It has a huge community of users so it will be easier to find support and it is used by some of the biggest marketplaces in the world so we don’t need to worry about NodeJS not being performant enough.

Python will be used in order to train machine learning models from our data. We chose python for this task because it is the most common language for machine learning. It has very performant libraries like numpy and scikit-learn that provide functionality for manipulating data and creating models that you cannot get in other languages like JavaScript and Java. Additionally, it is the most familiar language for us to use for machine learning because almost every machine learning course teaches ml using python.