If you use expiring (also known as “volatile”) keys in Redis, you may be in for a surprise when you attach a new Redis replica to your database: the key count on your replica may be significantly lower than the key count on your primary DB. This is especially common if you have a large number of volatile keys.
Is your Redis replica missing keys? Did you just lose data? The short answer is, “no”. However, it is helpful to understand why, exactly, your Redis replica is reporting a lower key count even though you haven’t lost any data. This involves two implementation details: how Redis expires keys and how a Redis sends your dataset to new replicas.
In Redis, volatile keys are not expunged from memory at the exact moment that they are set to expire. Instead, they are deleted via one of two methods:
When a Redis client performs a read or write operation on a key, the Redis server will first check to see if that key exists and has an expiration time. If it exists and has an expiration time in the past, Redis will immediately expunge the key from memory before handling the command.
To prevent volatile keys sticking around in memory forever if they are not accessed, Redis uses a simple passive algorithm to expire keys: Every 10 milliseconds, it grabs 100 random volatile keys and immediately expunges any keys whose expiration times have past. If 25 or more keys are expunged, Redis immediately grabs another 100 keys and begins again.
The second method above is important because it means that a significant number of
keys may be volatile keys that are already “expired” but that have not yet been
expunged from memory yet1. Redis will continue to include these
keys in the “keys” and “expires” counts in your INFO output until they are
expunged.
When a Redis replica attaches itself to a Redis server, the primary server creates an RDB snapshot of its dataset and sends it to the replica. And when Redis creates RDB snapshots, it does not include any keys with expiration dates in the past, even if they haven’t been expunged from memory yet.
When you attach a replica to a Redis instance, that replica is receiving a dataset that does not include the expired (but still extant) volatile keys from the primary database. This is the same reason that your key count may drop when restoring a Redis server from an RDB backup.
Special thanks goes to @mattsta for help tracking down the RDB behavior.
In practice, this number is usually much lower due to the fact that Redis is constantly looking for keys to delete. ↩