Struct CrossNodeGramReduction 

pub struct CrossNodeGramReduction { /* private fields */ }

Coordinator-side reduction: receives NodePartials from the fleet and folds them into the single global fixed-tree accumulator.

Receipt validation is per rank and in-sequence: rank r’s i-th accepted partial must be its i-th owned chunk. This makes the per-rank cursor in CrossNodeCheckpoint::received_per_rank a complete description of what has been received, which is what lets a dead rank resume from a bare ordinal. Cross-rank arrival order is unconstrained (the inner accumulator buffers out-of-order chunks), so slow nodes never block fast ones.

Implementations

impl CrossNodeGramReduction

pub fn new(partition: CrossNodePartition) -> Result<Self, String>

Fresh coordinator for the given partition.

pub fn partition(&self) -> CrossNodePartition

The shared partition (workers must be constructed with an equal one).

pub fn checkpoint(&self) -> CrossNodeCheckpoint

Serialize the full coordinator state. Resume-equals-straight-through is inherited bit-for-bit from the inner accumulator; the per-rank cursors resume receipt validation exactly where it stopped.

pub fn resume(state: CrossNodeCheckpoint) -> Result<Self, String>

Reconstruct a coordinator from a checkpoint, validating the cursor structure against the partition so corruption is rejected loudly.

pub fn finish(self) -> Result<Array2<f64>, String>

Finish the pass, returning the k×k border Gram. Errors if any rank’s sequence is incomplete. The result is a pure function of the row content and (n_rows, chunk_size) — identical bits for any node count, any arrival interleaving, and any checkpoint/resume history on either side.