v3.42.0
Changed
Section titled “Changed”- The node agent image is 76 MB instead of 1.2 GB. It no longer carries a CUDA runtime it never used, and it detects GPUs directly from the system. New nodes, and edge nodes on slow links in particular, are up much faster.
- Nodes pull AiQu images from a new registry. The join command a new node
gets, the Kubernetes manifests and the version check all use it. An
installation that needs a different registry can set it with the
AIQU_REGISTRYenvironment variable.
- System services could stay yellow on a newly joined node, because they still pointed at the retired registry. Nodes that had been running for a while hid the problem behind their local image cache.
- Creating or cloning a job with environment variables could fail with “Error adding job environment variables” when the optional delete-after-use flag was left out. Cloning such a job from the job list failed every time.
- The node agent could restart in a loop on startup, which showed as a node that never turned green, with nothing in its log pointing at the cause.
- The node tunnel could come up without routing any traffic.