Batching and matching

BATCH

Groups entities by count only. Each arrival is put into the block's own store; as soon as batch_size items are waiting, exactly that many are taken atomically (one synchronous store operation, so two arrivals at the same instant cannot both trigger a batch from overlapping items) and combined into one batch entity of type batched_entity_type. The members leave live_entities -- the batch entity stands for them in every KPI -- and carry in_batch until a SEPARATE splits the batch again.

SEPARATE

Splits a batch entity back into its batch_elems children. mode controls what the children's attributes look like afterwards:

Mode	Children get
Retain original entity values	Their own attributes, from before they were batched.
Take all representative values	Every attribute of the parent batch entity (identity/lifecycle fields excluded, so children keep distinct ids).
Take specific representative values	Only the attributes listed in attrs.

The parent is always destroyed at SEPARATE; its accumulated cost splits evenly across the children (see Costing).

MATCH

Matches number_to_match distinct entity types together (not Arena's attribute-based match -- a deliberate simpler semantics, documented rather than replicated). The arriving entity is always stored first, so it can never be dropped while waiting; the match is then taken atomically, oldest waiting entity of each type first. batch_option controls whether the matched entities are combined into one batch entity or forwarded separately.

DUPLICATE

Forwards the original entity unchanged, then forwards number_of_duplicates - 1 shallow copies, each a fresh entity with its own id -- unlike BATCH, this does not consume anything; it multiplies the flow.

When to use which

Need	Block
Group N arrivals of one flow into a unit load	BATCH / SEPARATE
Synchronize arrivals of different types (e.g. a part and its paperwork)	MATCH
Fan an entity out into several independent copies (e.g. broadcast to several downstream stages)	DUPLICATE