Skip to main content

​
Transaction-batched Active V3 text deltas benchmark

Measured 2026-07-31. The acceptance target passed: for an unpartitioned Active V3 text index, insert transactions changed from N FTS uploads and N manifest entries to one upload and one manifest entry for batches of 1, 10, 100, and 500. Search-result digests were identical before and after compaction and across all three revisions.

​
Revisions and protocol

  • Three independent runs per case, three warmups per run, and 25 measured samples per run: 75 measured samples per revision, batch size, and latency mode.
  • Release builds with rustc 1.97.1 (8bab26f4f 2026-07-14), LLVM 22.1.6, aarch64-apple-darwin.
  • Apple M4 Pro Mac16,7, 24 GiB RAM, Darwin 25.5.0. No benchmark revisions ran concurrently.
  • Fresh local-filesystem database, identical deterministic documents, configuration, queries, and cache state for every sample.
  • Modes were local storage and local storage with a synthetic 1 ms delay applied only to FTS blob uploads.
  • Compaction was held through the immediate search measurement. The database was then reopened with automatic compaction and observed until quiescent before the post-compaction search.
  • Median and p50 are the same nearest-rank observation. p95 uses index ceil((n - 1) * 0.95) over the 75 aggregate samples.
The complete aggregate metrics and all 24 independent-run summaries per revision are in text_transaction_batching_benchmark_results.json.

​
Transaction latency and throughput

Times are aggregate p50 / p95 milliseconds. Throughput is aggregate p50 inserts/second. Against the primary PR #47 baseline, local p50 transaction latency improved by 3.5x, 28.5x, and 84.7x for batches 10, 100, and 500. Under synthetic upload latency, the improvements were 3.5x, 29.5x, and 87.5x.

​
Primary structural and search comparison

Counts and bytes are p50; their p95 values were identical in every case. Search times are p50 / p95 milliseconds. The only apparent regression was local batch-1 immediate search: p50 increased by 31 microseconds (3.0%) and p95 by 8 microseconds (0.5%). Transaction latency and post-compaction search improved for that case, and the change is small relative to run dispersion, so it is classified as measurement noise rather than a batching regression.

​
Correctness and delete-only acceptance

Each batch had one digest across every measured sample, latency mode, revision, and immediate/post-compaction pair: The insert benchmark does not manufacture a delete-only batch. The production-linked interpreter_active_text_transactions_batch_by_destination contract separately asserts that delete-only destinations update root/state while producing zero uploads and zero manifest growth.