WIKIPEDIA · 5,032,104 DOCUMENTS
Search Benchmark, the Game
Five search modes, RLucene vs. Apache Lucene. Verified results before measured performance.
01 / QUERY PERFORMANCE
Query performance
Query time distribution Distribution of per-query medians over 12 rounds; μs
| Both batches | Validation | Details |
|---|
02 / METHODOLOGY
Traceable numbers. Consistent methods.
Each comparison uses the same query input, index and paired round. Absolute timings from separate batches are kept separate. Filtered results come from the full mixed workload; these groups were not rerun in isolation.
How to read the numbers
- Per-query time: the median of 12 rounds in this batch, including process communication, query parsing, full search and response.
- Relative delta: the median of (Rust / Java − 1) × 100% calculated for each paired round. It may differ from the ratio of the two displayed median times.
- Selected-query total: sum the selected queries within each round, then take the median. The aggregate delta uses paired round totals.
- Mean and P50/P90/P99 describe the distribution of per-query median times for the selection, rather than request tail latency in an online service.
- Faster and slower describe observed direction, without implying statistical significance. All regressions and reversals between batches are retained.
Per-query result equality
Before timing, query identity and mode are checked individually. COUNT verifies exact counts; TopN verifies returned counts, IDs, order, f32 score bits, totalHits and relation. Full outputs must be strictly equal.
For ordinary TopN, totalHits may be a lower bound. The timing response of 1 only acknowledges completion; it is not a document count. Full correctness verification runs separately.