On load-test day, the P99 for the order details endpoint jumped from 40ms to 900ms. The database was idle, the application reported no errors, and the slow query log was empty. Eventually, monitoring showed one Redis node pegged at 100% CPU while using only 20% memory—a single product-detail key was taking 80k QPS on a single instance.
1. A Hot Key Is Not a Big Key
People often conflate the two, but the solutions are completely opposite.
A big key means the value is too large : a Hash stuffed with hundreds of thousands of fields, and a single HGETALL saturates the network and memory. The fix is to split the data structure.
A hot key means access is too concentrated : the value may be only a few hundred bytes, but it gets read tens of thousands of times per second. The fix is to spread out the access.
The trickier part is this: Redis Cluster uses CRC16 of the key modulo 16384 to assign 16,384 slots, and a key always lands in the same slot on the same node .
So:
Adding nodes won't solve a hot key.…