Modern data platforms have become sprawling zoos — teams routinely run HDFS, YARN, ZooKeeper, Hive Metastore, HBase, Spark, Presto, and now Slurm or Kubeflow for GPUs, each with its own storage copy, security model, and failure surface. YTsaurus, open-sourced by Yandex under Apache 2.0 after 15 years in production, collapses that stack into a single distributed platform: one storage layer that serves batch ETL, sub-10ms KV lookups, ClickHouse and Spark analytics, and multi-node GPU training for LLMs — with unified access control, TLS, OAuth, and a Kubernetes operator for turnkey deployment. In this talk we tour the architecture (Cypress metadata, static and dynamic tables, GPU-aware scheduler), map each layer 1:1 to its Hadoop-stack equivalent, and show how the same cluster runs an exabyte-scale ETL pipeline, an interactive SQL workload via YQL, and LLM pre-training on shared hardware. You’ll leave with a concrete map of what YTsaurus would replace in your current stack — and a clear entry point if you want to evaluate it yourself.