Explain the internal working of Qdrant's sharding and replication?

Sharding splits a collection's data horizontally across multiple nodes so no single machine needs to hold the entire dataset, while replication keeps multiple copies of each shard on different nodes for fault tolerance — the two mechanisms work together to give a distributed Qdrant cluster both scale and resilience.

When a collection is created with a given shard_number, its points are distributed across that many logical shards (commonly by hashing each point's ID), and each shard is then replicated according to the configured replication_factor, spreading both the data volume and the read/write load across the cluster's nodes rather than concentrating it on one.

A query arriving at any node in the cluster is routed by a coordinator to the relevant shard replicas, which execute the search independently and return their local results, which the coordinator then merges into one final ranked answer — a process that's transparent to the client, which interacts with the cluster as if it were a single logical collection regardless of how many physical shards and replicas actually underlie it.