TLDR Tech / Data
Scaling Recommendation Systems with Request-Level Deduplication
Pinterest Engineering introduced request-level deduplication to efficiently scale their recommendation systems by sorting data by user + request ID in Apache Iceberg for massive compression to process and store the request-level data only once per unique request, using a separated context transformer with KV caching in ranking, and applying targeted fixes like SyncBatchNorm and user-level masking during training.
Read the full story