Inode exhaustion in flat cache directory

An image-resizing service that caches every derived thumbnail as a separate file under /var/cache/thumbs starts failing with ENOSPC on writes at 09:30, yet df -h on that volume shows only 60% bytes used. Resized images are tiny (a few KB) and the cache has accumulated tens of millions of them across a flat directory structure since the cache-cleanup cron was accidentally disabled in a refactor three weeks ago. A campaign this morning drove a surge of new unique image-variant requests. Beyond the write failures, directory listings and lookups on that path have become very slow. Triage and remediate.