- Fix: the tables were wider than the page in a 1440 px window, so Health and Power were out of sight. Text columns
(pod, node, namespace, model, names, messages) now share the free width and shrink down to their minimum with an
ellipsis while staying readable, and Health and Power come right after the identity columns on the GPUs and
Allocation views and before the VRAM columns in the Pods table and drawers. Column widths resized by hand are stored under a new key, since some
views changed order (#10). - Fix: when a scrape failed, the rows of the last good one stayed on the page under the error with the status still
counting the exporters, and nothing said they were old. The status now reads "stale: last good scrape HH:MM:SS",
the pages built on the scrape (Pods, GPUs, Idle & waste, Allocation) and the drawer tables are dimmed, and the
pages built on the pod list keep full contrast. When every exporter scrape fails, discovery runs again at the next
tick instead of scraping the same dead pods until its cache expires (#9). - Fix: discovery took the node name inside a pod's name for a GPU keyword, so on a node called like
gpu-*the
static control plane pods (kube-apiserver, etcd, scheduler, controller manager) were probed at every discovery pass.
The node name is now left out of the pod name before matching (#7). - Fix: on a cluster without exporters the whole cluster was listed again every 20 s (pods, services, nodes), also
when only a Pod or Node drawer was open. An empty discovery and a failed Prometheus search are now cached for 60 s
like a successful one, nodes are listed with the discovery instead of with every scrape, and the drawers poll only
for a pod that can use a GPU (or on a cluster known to have GPUs) and for a node that advertises one (#8).