A few weeks ago, Meesho announced open-sourcing its machine-learning platform. The news—first covered by AIM—quickly received praise from all over the world, especially the open source community, which urged the so-called unicorns to follow suit.
Recently, AIM caught up with Debdoot Mukherjee, the head of AI and data science at Meesho, to learn about the e-commerce company’s open source initiatives, why the company chose to give back to the community, its significance, and more.
At the NVIDIA AI summit in October, Mukherjee provided a deep dive into Meesho’s machine learning infrastructure. The ML architecture, which handles Meesho’s platform used by over 2 million sellers and 160 million plus consumers, was open sourced.
Meesho currently uses AI and ML for catalogue listing, pricing recommendations, ranking systems, fraud detection and user engagement.
“It is built to deliver unparalleled cost-efficiency,” said Mukherjee, referring to Meesho’s ML platform at the NVIDIA AI Summit last month. Interestingly, he also said that Meesho operates at the lowest server costs per order.
The M in Meesho Stands for Machine Learning
Mukherjee told AIM that Meesho plans to open source its feature store first, followed by more components in the future. A feature store is a centralised storage system for all data inputs the model requires to make predictions.
The store handles millions of requests per second, and 99% of these requests are handled in under 10 milliseconds. It is also said to serve two trillion features per day.
Further, Meesho said that it is looking to introduce feature groups on its ML platform. This essentially creates a class of similar features that can be generated, stored, and reused together. By handling these features as a single unit, significant compression is achieved due to shared redundancies.
Moreover, its model inference serves more than 500k requests per second in prediction. Meesho has also leveraged NVIDIA’s Triton Inference server, along with Kubernetes to scale its ML models, for cost optimisation and performance efficiency.
Mukherjee said they were able to drastically reduce the pod boot-up time from 6 to 7 minutes, to just 10 seconds.
Meesho uses TensorRT as the backend of the Triton Inference Server, which has reduced costs by 40%. It now takes $1.5 per 100M inferences, at a latency of 25ms. The platform used post-training quantisation techniques, which helped it achieve an eightfold reduction in costs and a fivefold improvement in latency.
Moreover, instead of upgrading to a more expensive GPU, Meesho increased the batch size on its existing NVIDIA L4 GPUs. All of these techniques led to a ninefold decrease in costs and a tenfold improvement in latency.
Mee-Sho Altruistic
Mukherjee believes that Meesho has built a platform with ‘fairly niche’ capabilities that aren’t available on open source markets yet. He said that their alternatives are mostly proprietary, and startups would have to pay a hefty price if they were to use them.
“If startups start building on top of our projects, hopefully, life will be simpler for them. They will be able to take large-scale machine learning workloads to production much faster and for much cheaper as well,” said Mukherjee.
By open-sourcing its ML model, Meesho will benefit from community feedback while allowing open source developers to build upon its work. Mukherjee said that the feedback regarding bugs and issues will help them make the architecture more robust.
“Whenever an open source project becomes successful, the community builds on top of it. They build capabilities and features, which can be potentially useful for us too,” said Mukherjee, who believes further development on the open source parts of Meesho’s ML infrastructure will be driven by contributions from the community.
Not a Zero-Sum Game for Indian Startups
When discussing open source in the Indian context, Zerodha deserves a mention. The company recently announced a $1M FOSS fund to invest in open source projects. In an interview with AIM a few months ago, Zerodha CTO Kailash Nadh said that almost every startup in India leverages the best of open source tech, but very little is being done to give back to the community.
Expressing a similar sentiment, Mukherjee acknowledged that Meesho has used several open source technologies in its tech stack, and now wants to give back. “We benefit quite a lot from open source. A lot of it [Meesho’s tech stack]is built on top of open source projects. So, it’s time to give back to the community,” he said.
While there’s room for more, Indian startups are certainly doing their bit to give back to the open source circle. With this move, Meesho has joined the group of Indian startups open-sourcing a part of their tech stack. The list includes CRED, Flipkart, Swiggy, PhonePe, Zomato, and Zerodha. Besides, many Indian startups have public repositories on their GitHub page.
Zomato recently open sourced its weather monitoring system, and design system Sushi. Similarly, CRED, which sells itself on design and user experience, has open sourced its neumorphic design system, NeoPop.
A public repository on Flipkart’s GitHub, with over 5,000 stars, called RecyclerList View, provides a high-performance list view component for React Native and web applications. One of PhonePe’s public repositories’, Mantis, is a security framework designed to automate the workflow of asset discovery and vulnerability scanning.
The post Why the Indian Open Source Community Loves Meesho Now More than Ever appeared first on Analytics India Magazine.
