Module parallel 

Intra-query parallelism under one process-wide thread budget.

A long read (an aggregate over a million records, an exact vector scan) can split its key range into parts and run them on scoped threads. Every query draws helper threads from one shared budget, sized to the server’s worker threads (RUSHDB_THREADS), so concurrent queries never run more threads than the host was given: a query that finds the budget spent runs its parts on its own thread. Parts are numbered and their results are returned in part order, so a caller that merges them in order gets the same answer as a sequential run whatever the timing.

Functions

budget

The helper-thread budget set_threads configured.

parts_for

How many parts a job of units should split into: one per helper the budget has plus the caller, and none smaller than min_units.

run_parts

Run part(0..parts) and return the results in part order. Parts run on the calling thread plus as many helper threads as the shared budget lends right now (at most parts - 1); with none free, they run one after another here. The first error (by part order) is returned.

set_threads

Set the helper-thread budget: the host’s worker threads less the one each query already runs on. Called once at startup; 0 = run every part on the calling thread.

split

Split 0..len into at most parts contiguous ranges of near-equal size, in order (no empty ranges).