Event language
UI language
<p><strong>INTRODUCTION:</strong></p><p>Deploying machine learning models in a production environment can be a challenging task, as it requires a balance between model complexity, scalability, and cost-efficiency. In this talk, we aim to share our experience and expertise in building <strong>robust, near-real-time machine learning pipelines</strong> that can handle high-volume, high-velocity, and high-variety data streams.</p><p>OBJECTIVE:</p><p>The main objective of this talk is to provide an in-depth understanding of the technical challenges and best practices involved in building near-real-time machine learning pipelines. We will focus on the following key aspects:</p><ul><li><p>Leveraging a <strong>distributed event-streaming platform like Kafka</strong> for data streaming in ML systems</p></li></ul><ul><li><p>Importance of <strong>pre-computation feature pipeline</strong> in the ML systems</p></li></ul><ul><li><p>Building a feature store that has 4 key attributes - scalable, highly available, handle bursts of reads/writes per second, schema safe for downstream data stitching</p></li></ul><ul><li><p>Asynchronous inference service for large request payloads </p></li></ul><ul><li><p>Implementing <strong>cost optimization techniques</strong> to build ML infrastructure</p></li></ul><p>LEARNING OUTCOMES:</p><p>The participants can expect to learn more about the following during the course of the talk:</p><ul><li><p>The challenges and <strong>best practices of deploying machine learning models</strong> in a production environment</p></li></ul><ul><li><p>How to <strong>leverage Kafka to build a pre-computation layer</strong> for feature generation</p></li></ul><ul><li><p>Challenges in building a feature store for scale</p></li></ul><ul><li><p>Managing <strong>multiple read/writes into a feature sink </strong>for highly available downstream applications</p></li></ul><ul><li><p><strong>Evaluating different ML API serving</strong> open source platforms like Cortex, and Seldon core versus managed services like Amazon Sagemaker</p></li></ul><ul><li><p>Applying cost optimization techniques - <strong>scalable inference at up to 85% lower cost per inference</strong></p></li></ul><ul><li><p>How to <strong>build a queryable data lineage management layer</strong> over the inference events</p></li></ul><ul><li><p>Log inference outcomes like your life depended on it - and monitor events like a hawk</p></li></ul>
