Could you DM me a curl request showing the issue? I wonder if your cluster ran out of RAM, so the first request is swapping content from disk to RAM