Arek Borucki describes the Hub's metadata and application architecture, separating MongoDB records from model artifacts in object storageObject storage keeps data as independently addressed objects with payloads and metadata in a flat, scalable namespace.. This allows storage, metadata and compute to scale independently. The talk focuses on tail latencyTail latency measures the slowest fraction of requests, commonly expressed with high percentiles such as p95, p99 or p99.9. and keeping the user experience simple as the catalogue expands.
Arek Borucki explains how model names are tokenized at insertion and stored in a denormalized read collection. The team replaced regular-expression search with Lucene-backed Atlas Search as the dataset grew, while maintaining ranking based on a changing popularity score.
Arek Borucki describes distributing suitable reads and aggregations to replicasDatabase replication maintains copies of database information on more than one node according to a defined update process., reserving the primary for writes and consistency-sensitive operations, and isolating heavy reporting on a hidden node. Sharding is presented as a next step rather than an already completed migration.
Arek Borucki distinguishes application-pod scaling from adding underlying cluster capacity. The talk also describes plans to move toward application-level demand signals rather than relying only on resource utilization. Reported scale figures belong to this presentation, not a continuously updated service benchmark.
Watch on YouTube




