Cloud migration improves team productivity and data science model visibilityTelecommunications Industry
This transformational journey entailed a set of data migration and data science model migration projects. The data was migrated from on-premise SQL databases to the cloud based Snowflake data sources. At the same time, the on-premise manually refreshed data science models needed to be re-written in Azure ML pipelines.
FOYI was involved in migrating the data science models to Azure ML on the cloud.
Challenge
The telecom customer initially proposed a project scope with the task to migrate 6 on-premise production models to the cloud in 6 months. Furthermore, it was required that these models be enhanced by performing feature engineering and where needed have new features built.
There were 3 key challenges in this objective identified by the business and FOYI.
- The cloud architecture i.e. Azure ML was evolving rapidly.
This meant that what worked in Week 1 of the project will change significantly by Week 8. FOYI demonstrated that the ability to upload and use a model object as a pickle file at the time of Request for Proposal(RFP) was not available by the time the project scope is being discussed. - The underlying data sources were also getting migrated at the same time.
This meant that training data for the models was on-premise and prediction data was on the cloud. FOYI brought out the challenge of deploying trained models as the new features were not available in the same format as the old ones in the training data. - Four out of the six models in question were used as weekly inputs to marketing teams and hence cannot be stopped.
This meant that migrating these models must be done only when the entire end-to-end process of migration was well known.
Approach
As a first step, FOYI proposed that a pilot project be undertaken with the intention of just one project be migrated as a lift and shift project. The proposed objective was to relocate the on-premise model to the cloud without any enhancement but with changes where the new features be replaced with existing ones where they are not available in the exact format. This will help in estimating the effort for the project more accurately.
This proposal led to the customer revising the objective to a pilot project with 2 models that needed monthly refreshes.
Outcome
The 3 key deliverables of this project and their benefits are as follows.
1. Migration of 2 production models to the cloud
Benefit: Automation of the monthly refresh process means more productivity for the team
2. Identify & mitigate the impact of the new data sources on the model features.
Benefit: Feature stores were updated with the substitute features thus reducing the time to find them for other models using them.
3. Detailed documentation of all the challenges of model migration and the solutions/workarounds thereof.
Benefit: No dependency on external consultants for migrating the rest of the models and therefore low cost of migration project.