Talks

How we solved the problem of making hotfix changes to ETL pipelines on Apache Spark in hundreds of existing processes without changing their code.

Our Trino storage hit the performance ceiling of a single Ceph cluster — so we started spreading every table across several clusters at once, hiding all the sharding logic in the HAProxy sidecars on our compute nodes, without adding a single new component to the architecture. Reads sped up from 20 to 60–80 GB/s, and GET latency dropped from minutes to 1–2 seconds.

The talk is about practical experience in optimizing inference and ML-serving based on GPUStack in the production environment of the corporate AI Portal.