v2 vs v1 — Performance (v2)

The measured v2-vs-v1 comparison — prediction accuracy and latency, cold-start warm-up, scaling, memory/disk footprint, write throughput, and a direct Elasticsearch head-to-head — now lives in one place so every number has a single home and can't drift between copies:

→ Benchmarks → Query, predict & write speed

The short version from that page:

  • Accuracy (10M rows). v2 leads the mature v1 engine on processor and trails it on acceptor and glCode by 2.5–4.5 points — all within the noise band of the ~511-row hold-out, so no target separates the engines. See the performance benchmarks for the per-target numbers.
  • Latency. Interactive — roughly a tenth to a half a second per predict at 10M rows — with v2 cold-start markedly faster than v1 on every target (memory-mapped columnar reads).
  • Footprint & scale. Predict latency grows sub-linearly with data on the tested targets (interactive — sub-second per predict — through 10M rows), a leaner memory-mapped columnar footprint, and incremental writes (no reindex step).

These figures are measurements of specific builds on specific corpora, not a performance guarantee — read them as "where rep2 is on these targets".