Clustering Interview Questions

Cluster hash slots are 16384; a key hashes to one slot and that slot lives on one master. MOVED redirects the client when the slot moved. Clients must be cluster-aware or they will keep hitting the wrong node.

Hash tags {userId} in the key force several keys into the same slot so MULTI or Lua can touch them. Overusing one tag such as just {user} creates a hot slot. Tags are for related keys, not a global namespace.

Cross-slot errors happen when a transaction or Lua script touches keys that hash to different slots. The fix is a hash tag, not a retry loop. MGET of random keys is also cross-slot unless they share a tag.

ASK redirect is temporary during slot migration: the client should ASK the destination for that one command, then resume. Treating ASK like MOVED forever is wrong. Resharding is when you see ASK.

Replica failover in Cluster elects a replica of a down master and moves those slots to it. Clients see MOVED. replica-priority and replica hang detection matter. Sentinel is a different HA story for standalone Redis.

Cluster vs Sentinel: Sentinel watches a primary-replica set and failovers a single keyspace. Cluster shards 16384 slots across masters. Do not run Sentinel in front of Cluster nodes as if they were standalone.

Resharding moves hash slots between masters with redis-cli --cluster reshard or Cluster Manager. Keys in those slots stream to the new master. Watch ASK and MOVED rates and avoid heavy KEYS during the move.

READONLY lets a client read from a replica in Cluster; data can be stale by replication lag. READWRITE returns you to the master. Serving GET from replicas without READONLY fails.

cluster_state fail plus cluster-require-full-coverage means if too many masters are unreachable the cluster refuses writes. That is safer than serving a split brain. Setting coverage to no is a conscious availability trade.

Adding a node starts with CLUSTER MEET, then assign slots or CLUSTER REPLICATE to make it a replica. Empty masters with no slots still participate in gossip. Forgetting CLUSTER MEET is why a node stays isolated.

Hot slots: one slot getting most writes, often a celebrity hash tag or low-cardinality key, cannot be split below one slot. The fix is a better key design, not more hardware. Empty neighboring slots are normal; hotness is about key distribution.

Cluster bus port is the Redis port plus 10000 by default and must be open for gossip. Opening only 6379 and then wondering why failover never happens is a classic ops miss. TLS on the bus is a separate setting.

MULTI/EXEC in Cluster is still single-slot. There is no distributed transaction across slots. If you need cross-entity atomicity, you tagged those keys together or you use an external workflow.

Migrating from Sentinel to Cluster is a reshard of a previously unsharded dataset: dump and reload with hash tags planned. Applications must switch to cluster clients that handle MOVED. Dual-running both modes on one port does not work.

Failure of a replica is cheap; failure of a master with no replica is slot downtime. Every master should have at least one replica. replica-validity-factor stops stale replicas from winning failover.

CLUSTER SLOTS versus CLUSTER NODES: clients cache slot maps from SLOTS. Stale maps cause extra MOVED. After reshard, expect a burst of redirects; that is health, not an outage, unless they never converge.

Cross-slot Lua: even EVAL must list keys that map to one slot. Passing extra keys in ARGV to sneak around KEYS is unsupported and breaks under Cluster. Hash-tag the keys the script needs.

Network partition: Cluster uses a majority of masters to elect. A minority side stops taking writes if it cannot reach quorum. Clients on the minority must fail, not silently write. That is the CAP answer for Redis Cluster.

replica-announce-ip and cluster-announce-ip exist because Docker hosts often gossip unreachable addresses. Failover then looks fine in CLUSTER NODES and dead from apps. Cluster is unforgiving of wrong advertised IPs.

Slot migration stuck: a key that keeps being written during migrate, or a huge key, can stall a slot. UNLINK big keys first. CLUSTER GETKEYSINSLOT helps you see what is in the way.