DeepSeek V4 debuts production-grade Agent sandbox DSec: single cluster scheduling hundreds of thousands of concurrent instances, unifying four heterogeneous bases.

ME News, April 24 (UTC+8): According to Dongcha Beating monitoring, the DeepSeek V4 technical report publicly disclosed for the first time the core infrastructure supporting Agent post-training and massive evaluations— a production-grade elastic compute sandbox, DSec (DeepSeek Elastic Compute).

Current large-model reinforcement learning requires an extremely large code trial-and-error environment. The report reveals that in real-world production, a single DSec cluster can simultaneously schedule hundreds of thousands of concurrent sandboxes.

The system is written in Rust. At the underlying layer, it connects to the self-developed 3FS distributed file system, and breaks the performance bottleneck caused by massive sandbox cold starts through hierarchical on-demand loading.

In terms of developer experience, DSec uses a set of Python SDKs to unify four execution substrates: function calls, containers, micro virtual machines, and full virtual machines. When switching between them, only one parameter needs to be changed.

To address the task preemption problem common in compute clusters, DSec introduces a global trajectory log. When a task resumes, the system directly “fast-forwards” to replay the cached command execution results, achieving extremely fast checkpoint resumption while avoiding non-idempotent errors caused by repeated execution.

(Source: BlockBeats)