Add scripts to inject container mount points during the CRIU restore#2994
Add scripts to inject container mount points during the CRIU restore#2994wangixt wants to merge 1 commit intocheckpoint-restore:criu-devfrom
Conversation
…phase
When managing inference containers (e.g. vLLM) on Kubernetes, restore flows
often need extra mount points for model weights so that model switching can
happen without rebuilding container images.
This commit introduces:
- scripts/criu-move-mount.c:
- a CRIU action helper that parses CRIU_ADD_MOUNTS and performs bind/rbind
mounts into the restored container mount namespace during pre-resume.
- scripts/runc-action-add-mounts.sh:
- an action-script wrapper that compiles criu-move-mount.c at runtime and
executes it during CRIU restore.
The solution enables best-effort mount injection at restore time, reducing
operational friction for dynamic model weight switching.
|
@wangixt Why do you want these scripts to be in the CRIU repository? |
Thanks for raising this — that’s a fair question. Our motivation for proposing these scripts in the CRIU repo is that they solve a generic CRIU restore-time workflow, not a vLLM-specific one:
That said, we understand scope concerns.
We’re flexible on placement; the main goal is to provide a reusable restore-time mount injection pattern for CRIU users. @rst0git |
|
@wangixt Have you considered using the "prestart" OCI runtime hook or just modifying the container config to include the additional mounts?
Using a CRIU action-script to inject mounts doesn't sound like a reliable approach for doing this. |
When managing inference containers (e.g. vLLM) on Kubernetes, restore flows often need extra mount points for model weights so that model switching can happen without rebuilding container images.
This commit introduces:
The solution enables best-effort mount injection at restore time, reducing operational friction for dynamic model weight switching.