DeepSeek V4 Unveils Production-Grade Agent Sandbox DSec: Managing Hundreds of Thousands of Concurrent Tasks in a Single Cluster

According to monitoring by Dongcha Beating, the technical report for DeepSeek V4 has publicly revealed the core infrastructure supporting Agent post-training and massive evaluations, the production-grade elastic computing sandbox DSec (DeepSeek Elastic Compute). Currently, large model reinforcement learning requires an extremely vast code trial-and-error environment. The report discloses that in actual production, a single DSec cluster can simultaneously manage hundreds of thousands of concurrent sandboxes. The system is written in Rust and interfaces with the self-developed 3FS distributed file system, breaking the performance bottleneck of cold starts for massive sandboxes through hierarchical on-demand loading. In terms of developer experience, DSec unifies four execution bases—function calls, containers, micro virtual machines, and full virtual machines—using a single Python SDK, requiring only a parameter change for switching. To address the common issue of task preemption in computing clusters, DSec introduces a global trajectory log: when a task is resumed, the system directly ‘fast-forwards’ to replay the cached command execution results, achieving rapid breakpoint continuation while avoiding non-idempotent errors caused by repeated executions.