Advertising is one of Yandex's most highly loaded services. It handles up to 700K RPS and works with billions of ads.

Hundreds of ML models are responsible for the quality of ad ranking. The ML development department continuously trains new, more advanced models.

One of the complex problems that significantly impacts the metrics of online experiments is the discrepancy between training and serving data (training-serving skew). In Feature Store, we are developing a meta-layer. It allows ML developers to easily construct and reuse features for models in the context of low latency (real-time model serving) and high throughput (model retraining on the cluster).

What tasks await you

Integration with Search and Advertising services Feature Store has already become the standard for deploying neural networks in Advertising, and you will help us further integrate into various services, including Search and Advertising, as well as adapt to the needs of new model classes (generative transformers). You will solve various tasks: from optimizing computations in C++ to developing Python interfaces for users and integrating with internal services (YQL, Model Registry, model services). Your knowledge will help improve the user experience, speed up the launch of new models, and save resources.

Learn more about backend at Yandex — in the channel Yandex for Backend

We expect you to

• Have a good command of C++
• Have worked with high-load systems
• Know classical algorithms and data structures
• Have experience working with Python

Will be a plus

• Are familiar with the MapReduce concept and distributed computing engines (Hive, Spark, YT)