Parallel compounding occurs when independent components execute concurrently but total latency remains governed by the slowest component. Partitioning work enables lower latency, while increasing fan-out raises exposure to tail latency. section 2.4
MapReduce
MapReduce parallelizes its map phase because each output depends only on one input partition. The reduce phase gathers intermediate values, so job latency follows the slowest map task and becomes more variable as fan-out increases. section 2.4
Use parallel compounding deliberately:
- Partition work only when tasks are independent. section 2.4
- Measure latency variability, not only average component speed. section 2.4
- Distinguish parallel compounding from quorum compounding, which waits for a majority rather than every component. section 2.4
