Merge rules¶

`MergeRule`¶

Field	Meaning
`terminal`	the result shape, used in messages
`combine`	takes the arms’ partials in arm order (leading rows first) and returns the unsplit answer
`mergeable`	the aggregate kinds this shape can combine; consulted by `check_split`, default-deny
`failed`	optional predicate; a partial it accepts is returned as the merged result
`refusal`	when set, the shape cannot be split at all

MergeRule.merge(partials) returns the single partial when only one arm ran, so a degenerate share runs exactly the unsplit path.

Aggregate kinds¶

Set	Kinds	Condition
`ADDITIVE`	`sum`, `count`	partials add
`DEFERRED_MEAN`	`mean`	the partial carries sum and count separately; divide after
`BOUNDARY_STATS`	`min`, `max`, `first`, `last`	decided at range edges; the arms’ order defines first/last

AGGREGATE_REFUSALS holds the reason given for mean, n_unique, median, std and var. Any other kind outside mergeable gets a generic refusal that still names the aggregate and the terminal.

Helpers¶

Function	Combines
`merge_totals(partials, names, finalize=None)`	named totals by addition, then `finalize`
`merge_ordered_groups(partials, key=, fold=)`	per-group records of rows sorted on the key; folds the boundary group; refuses unsorted input
`merge_keyed_groups(partials, key=, fold=)`	per-key records with no ordering; result key-ascending
`merge_top_k(partials, limit=, key=, descending=True)`	the arms’ k-lists, stable re-rank
`merge_sorted_runs(partials, key=, descending=False)`	two sorted runs, earlier range first on ties

`check_split(rule, aggregates)`¶

aggregates are (output name, aggregate kind) pairs. The refusal names the output, so the caller knows what to change. It is separate from execution so a planner can ask whether a query is splittable before committing anything to either engine.