DeepSeek Elastic Compute (DSec)(arxiv.org)
315 points by shenli3514 17 days ago | 109 comments
tl;dr: DeepSeek's production sandbox platform, DSec, provides an elastic execution environment for large-scale LLM agentic training and RL workloads, exposing FnCall, container, microVM, and full-VM backends through a unified SDK backed by their 3FS distributed filesystem. It's co-designed with their RL framework to decouple stateful rollouts from preemptible GPU training. A single ~160-node deployment handles ~3 million sandboxes/day, 380,000+ concurrent sandboxes, and sustains 5,000+ sandbox creations per second.
HN Discussion:
  • ↑Impressed by the scale and technical numbers achieved on modest hardware
  • •Speculation about unpredictable workload allocation challenges in such systems
  • •Comparison to similar efforts by other companies like Google's ax
  • ↓Skepticism framing the system as just serverless rebranded
  • ↑Admiration for DeepSeek's innovation relative to better-funded competitors