Module shard_split 

Shard splitting: vector-aware and graph-aware partitioning.

When a shard becomes overloaded, it can be split into two shards. The splitting strategy is engine-aware:

• Vector-aware: split by collection + partition key. Each resulting shard holds a subset of the collection’s vectors with its own HNSW index. Cross-shard k-NN queries use scatter-gather with result merging on the Control Plane.

• Graph-aware: balanced partitioning minimizing cross-shard edges. Uses a greedy heuristic (BFS-based community detection) rather than full METIS for practical performance. Cross-shard traversals use existing scatter-gather infrastructure with ghost edges.

• Speculative prefetch: when scatter-gather dispatches to multiple shards, co-located vector and graph shards for the same tenant are prefetched together to reduce round-trip latency.

Structs

SplitPlan

A shard split plan: which vectors/documents move to the new shard.

Enums

SplitStrategy

Splitting strategy.

Functions

plan_graph_split

Plan a graph-aware shard split.

plan_vector_split

Plan a vector-aware shard split.

speculative_prefetch_shards

Speculative prefetch plan for cross-shard scatter-gather queries.