[tonyalaribe]

After #432 cut the indexes a warm 7-day errors list (6297304f) searches from 273 to 60, the query still spent ~3 s in the tantivy fan-out. It did not scale with hits (800 vs 200 hits both ~3 s; a window narrowed to today's small indexes: 32 ms) nor with the window range filter (day-aligned window still 2.4 s): it scales with the size of the day indexes in the window.

tantivy opens a fast-field column by reading it whole, and BundleDirectory::read assembled every range — hundreds of MB per day index, per query — through flat_map(|b| b.iter().copied().collect_vec()) with a vector per 64 KB block. It now copies each block's span with extend_from_slice.

Test: proptest a_bundle_range_reads_back_exactly (arbitrary ranges over block edges, cold and cached); red when the start offset is ignored.

Local signoff: make ci-signoff fmt/test/pg-smoke/e2e/clippy all pass (attested).

🤖 Generated with Claude Code

[monoscope-bot]

Monoscope production impact · Coverage unknown

Project 87576849-4941-49d3-a15d-680fef88a1a8 · reviewed 74a7e6791be964869f89b78bde03562d161d1ff4 · advisory

Evidence window (UTC): 2026-10-18 20:40:58.04257947 UTC → 2026-10-25 20:40:58.04257947 UTC

Coverage: Open Monoscope for evidence and coverage details.

Review history and settings

[claude]

Review: LGTM.

• Correctness: The new loop matches the old semantics. Only the first block is skipped, and the skip is clamped to the block length. The copy is capped at range.len() - out.len(), so it can't underflow. The Vec::with_capacity(range.len()) allocation is exact.
• Performance: Dropping the per-64KB Vec and the byte-wise flat_map/skip/take is the right fix. It becomes a memcpy per block.
• Tests: The proptest covers block edges, cold reads and cached reads. Two nits:
  • hi is always < len, so a range ending exactly at EOF (hi == len) is never generated. Use a.min(b)..=a.max(b) clamped to len, or % (len + 1) for the upper bound, to cover it.
  • (a.max(b) % len).max(a.min(b) % len) is hard to read. Compute lo and hi from a.min(b) and a.max(b), each reduced mod len + 1, then sort them.
• Not checked: If any block is None (it shouldn't be after the fill loop), flatten silently yields a short read, same as before. An ensure!/debug assert on out.len() == range.len() would turn that into a loud error instead of corrupt data.

No security concerns.