Question

Your write path uses a horizontally sharded datastore (data routed by `tenant_id` hash to 64 shards). During a product launch, one tenant goes viral. Shard 23's CPU and write latency spike to 10x the fleet median, its queue depth climbs, and a slice of traffic — all for that one big tenant — sees timeouts, while the other 63 shards sit nearly idle. Adding more shards in the past didn't help similar events. No deploy; pure traffic. Triage and mitigate.