Failing disk tail latency

p99 read latency on a stateful storage service jumped from 8ms to 600ms an hour ago, but p50 is unchanged and overall throughput is flat. The spike is isolated to one of twelve storage nodes: on that node, iostat shows %util near 100% with await in the hundreds of milliseconds and a growing queue depth, and dmesg has a handful of 'I/O error' / sector-remap lines. No deploy went out; traffic is normal. Requests that happen to hit that node's shards see huge tail latency. How do you triage, restore good tail latency now, and handle the failing hardware durably?