DRA Device Binding Conditions

This KEP introduces BindingConditions, a mechanism that defers Pod binding until external resources (e.g., fabric-attached GPUs, FPGAs) are confirmed ready. Unlike current scheduling, which assumes immediate readiness, this feature prevents premature binding and improves reliability for asynchronous or failure-prone resources. While the original motivation for this KEP was to support fabric-attached devices in composable disaggregated infrastructure, the proposed mechanism is designed to be broadly applicable. BindingConditions enable the scheduler to defer Pod binding until external resources (such as attachable devices, remote accelerators, or programmable hardware) are confirmed to be ready. This allows for more deterministic scheduling and avoids premature binding that could lead to Pod failures or manual intervention. The mechanism is not tied to any specific hardware model or infrastructure. It can support a wide range of use cases, including: - Fabric-attached GPUs that require dynamic attachment via PCIe or CXL switches - FPGAs that need time-consuming reprogramming before use If device preparation fails (e.g., due to contention or hardware error), the scheduler may clear the allocation and allow the Pod to be reconsidered for scheduling. This fallback behavior is part of error handling, but the primary goal is to enable more robust and predictable scheduling by explicitly modeling readiness before binding.