Context

With Udaan’s fast expansion in B2B e-comerce, our databases have to be highly scalable, resilient and performant. The choice of DB is one the most critical decision before we start building the solution. For our NoSQL usecases we chose azure CosmosDB which was distributed and easily scalable as well as resilient with failiures and retries.
But like any powerful tool, with great power comes great responsibility. And with that, thus came a huge bill as the scale increased.

Throughout our journey, we focused on two pillars - cost and stabilty. Over the last one and a half years, we have used several optimisations, some specific to cosmos like throughput model, and some generic NoSQL like tackling hot partition, effcective choices of partition key, indexes and composite indexes. We spent our fair share of time and effort to make frameworks, intelligent recommendation systems and tools to handle wide variety of use cases like daily cronjobs with heavy scales.
In the light of these efforts, not only have we managed to reduce more than 75% of the cost, we have also built a system to stabilise cost and spikes that comes with heavy and unpredictable loads.

Content of the talk

1. About Us and the problem we faced
2. Exploring challenges of NoSQL scaling in general as well as specific to cosmos
3. Stategies we used to optimize the db and queries for several types of usecases
4. Learnings
5. Building frameworks for smart recomendations and alerting for stabilizing of fluctuating costs with dynamic scale
6. Q&A