DeepSeek Elastic Compute (DSec)
- Source
- arxiv.org
- Author
- shenli3514
- Date
DeepSeek has described DSec, short for DeepSeek Elastic Compute, in a new arXiv paper. The system is the sandbox layer its agentic reinforcement learning runs on, and the paper reports it sustaining more than 380,000 concurrent sandboxes in production. It spans three isolation backends: containers, microVMs and full virtual machines. According to the summary, DSec decouples stateful agent rollouts from preemptible GPU training and coordinates the two so that reward hacking is reduced. That separation matters because agent rollouts hold state for long periods while training jobs can be interrupted, and coupling them wastes either GPUs or rollouts. Few labs have published infrastructure detail at this scale, so the paper is a useful reference for anyone sizing an environment fleet for agent training. The paper drew about 200 points on Hacker News within a day.
Context
Long-running training needs to preserve unfinished work when its GPU job is interrupted. DeepSeek's DSec infrastructure addresses this by running each agent's and a worker container that manages it outside the preemptible GPU pool. According to the technical report, these two components jointly hold the rollout state, so a preempted GPU job can reconnect and continue without rebuilding execution by replaying a command log.
When a job is preempted, the training framework also sends pause requests to its sandboxes. For containers, the process tree is frozen and memory is reclaimed while execution state is preserved; for microVMs, memory and execution state are saved in a snapshot and the running process is terminated. A later request to a paused sandbox resumes it. The paper describes this as a way to release resources while a job is interrupted without losing the environments' state.
Checking sign-in…
Loading comments…