Manticore Search 24.4.0 introduces parallel RT disk chunk merging and N-way merge jobs, replacing the old serial two-chunk-at-a-time compaction model. Two new settings control the behavior: `parallel_chunk_merges` (concurrent merge jobs) and `merge_chunks_per_job` (chunks per job). Benchmarks on 10M documents show compaction time dropping from 3:23 to 1:31 with tuned settings — a 55% total time reduction and ~90% shorter merge tail after ingest. For columnar storage, the improvement can reach 6.8x. Ingest throughput remains unaffected; the gain is entirely in post-load compaction speed. Optimal settings clustered around parallel_chunk_merges=4–5 and merge_chunks_per_job=4–5.
Table of contents
Why this mattersBefore: one merge job, two chunks at a timeAfter: parallel merges plus larger merge jobsWhat changed in practiceWhat about the new defaults?Broader benchmark results: row-wise and columnar storageHow to think about the two settingsTakeawaySort: