parallelization

Split the task into independent units with stable inputs, run each branch concurrently, then collect and merge outputs. Set per-branch timeouts, retries, and normalize outputs to a shared schema before merging with deterministic code.

Sectioning assigns each branch a different slice of work, such as one document chunk per branch. Voting runs multiple branches on the same input so several judges can score or rank a candidate answer for ensemble reliability.

Avoid it when branches need shared evolving context, when one branch's output changes another, or when merge logic is complex. Rate limits can also erase concurrency gains, and coordination overhead may exceed the latency saved.

A single long-tail branch dominates total latency, and cost grows linearly with branch count. Set per-branch timeouts, cap concurrency to fit rate limits, and benchmark against a serial baseline before scaling branches.

Define a retry and partial-failure policy before running branches, and log per-branch failures explicitly. Decide whether the merge step tolerates missing outputs or requires all branches to succeed, and never silently ignore failed branches.