Remote Code Execution

Execution infrastructure for trusted functions, generated code and long-running workloads.

Problem
Generated or customer-specific workflows needed controlled execution for long-running jobs, browser/session work and model workloads without reducing everything to trusted static functions.
My role
I forked and re-architected Alibaba Open Sandbox around execution infrastructure, security and product architecture.
Implementation
Separate trusted Lambda and dynamic Sandbox product modes; add nonblocking command lifecycle, TTL keepalive, pod-to-pod routing, node affinity, warm pods and on-prem trusted posture. The design handles workload examples such as 50MB inputs, 15 PDFs and ML jobs. Warm pods are assigned when already running, and warm nodes cache images through DaemonSets, balloon pods and Karpenter.
Engineering choices
Lambda means static first-party code with frozen dependencies, locked egress and SDK/internal gateway access. Sandbox means dynamic generated code with runtime dependencies and configurable network. On-prem execution uses the trusted Lambda posture: static first-party code, frozen dependencies, locked egress and SDK/internal gateway access.

Stack and domains