Deploying ML Pipelines for B2B
Setup of the ML pipeline in the cloud is important in B2B space for the following reason.
- For each customer, Model retraining is expected from time to time
- Model deployment has to happen for multiple customers.
If this setup does not happen, each time model training happens on your laptop, there would be no time for us to do any things. Solution is to give power to the Devops
Model Serving:
Model is built and stored in pickle or joblib file. API is built using a web application framework such as Flask, Django, Fastapi etc. This api has to be setup in the cloud. There are different ways to do it.
- VM: Setup the api in the VM. The scaling has to be taken care by the team.
- PAAS: Setup the api in the PAAS such as Azure app service and Aws Elastic bean. Scaling can be changed with the push of buttons.
- FAAS: Setup the api in the FAAS such as Aws Lambda and Azure function which takes care of scaling the system
- Docker: Setup the api in the docker image and build the kubernetes architecture to scale it. Docker is the created to remove the unnecessary burden of setting the system in different environments.
Setup Model Pipeline
- Shell Scripts: Shell scripts are created and each step from extracting the data to model building is executed sequentially. Shell scripts have to build from scratch. Bring best practise requires significant efforts.
- ETL Pipeline: Model building code can added into the ETL pipeline such as Azure Data Factory, Aws Glue. Pipeline templates can be downloaded and deployed. This helps in scaling to multiple customer easily
- ML Pipeline: Aws Sager Maker and Azure ML Studio make it so easy that model building and deploying is feel like happening in ipython notebook in my laptop.
- Kuberflow Tools: Kuberflow pipeline is built for End to End Pipeline. Difference from the above is that each module is the docker and Kuberflow is build top of kubernetes.
The next fews years, we will see an increase in huge number of tools in MLops. Usefull reference are as follows
The next article will talk about the job opportunities that are available, how ML is evolving and what kind of companies are best suited. , Please click here for next article
Please click here to go to the first article
