Scaling and distributed search concepts

Scaling and distributed search concepts explain how ANN workloads grow beyond one machine: sharding, replication, routing, fan-out, merges, skew, and disaggregated memory.
Created: Updated: 2 min read

These are the topics in this section, each on its own page with a stable path you can bookmark or share.

What topics are covered in this section?

Scaling and distributed search concepts explain how ANN workloads grow beyond one machine: sharding, replication, routing, fan-out, merges, skew, and disaggregated memory.

What changes when one graph is not enough?

Sharding splits a collection so each piece owns a local index; replication copies shards for availability and read scale. Random versus cluster-based sharding changes whether similar vectors co-locate. Routing indexes and meta-indexes steer queries; fan-out and top-k merge combine shard results; recall decomposition explains errors in that merge. Load skew, fault tolerance, RDMA, disaggregated memory, distributed graph traversal, and network-versus-compute trade-offs appear at cluster scale.

Distributed ANN is as much systems design as algorithm design.

How does Weaviate scale HNSW?

Weaviate shards for dataset size and replicates for HA and QPS; each shard carries its own HNSW (or other) index. Queries fan out and merge. Async replication and repair-on-read keep object replicas aligned – related to distributed graph repair in the frontier glossary. Multi-tenancy often maps tenants to shards so isolation stays sharp. HFresh reduces RAM pressure that forces early sharding.

Part VII’s multi-machine chapter pairs with these terms.

Start with sharding, replication, fan-out, and top-k merge, then load skew and fault tolerance. Continue into Part VII and distributed graph repair when connectivity across nodes is the concern.

This section is ANN at cluster scale. Next, read “What is sharding in a search index?” or Part VII’s multi-machine chapter, then replication pages when availability is the goal.