[hodaaaaaaaaaa]

Problem

A new worker on a node without the worker image pulls it first, several GB that take minutes, and a cold job pays it for its trainer and its sampler.

Change

The base bundle adds the open-rl-worker-prepull DaemonSet, which pulls the worker image onto every GPU node with openrl.io/enabled=true at deploy time and only sleeps. The kustomize images transformer rewrites its image like every other OpenRL image, so a release that bumps the tag re-pulls on every node. The kind overlays drop it, since their workers pull from a cluster-local registry.

Based on droot's prototype in #343.

Results

Measured on GKE with eight GSM8K GRPO jobs launched together (gemma-4-E2B and Qwen3-0.6B, LoRA and full fine-tuning, A100 80GB and L4), with the worker image removed from every node before each run, two runs each:

• A cold worker image pull took about 3 minutes per worker.
• With the DaemonSet rolled out first, every job reached its first training step 330–400 s sooner, and the whole workload took 749 s instead of 1111 s.