[tonyalaribe]

Observability only. After #426 a cold 7-day errors list on 6297304f reads ~1.1 GB of bundle ranges for 800 hits (~1.4 MB/hit vs ~255 KB/hit on the full path). Whether those bytes are fast-field columns, term data or the doc store decides the next fix (cache them on disk vs stop reading them).

Adds range_read_{fast,terms,store}_bytes to timefusion_stats (tantivy component); the remainder of range_read_bytes is other files. BundleSource::read_file is a provided method, so only the object-store source changes.

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 2bede6ab4f2ef16d320c2bf385677f87db5b54b2 · advisory

Evidence window (UTC): 2026-10-18 15:33:07.387265537 UTC → 2026-10-25 15:33:07.387265537 UTC

Coverage: Open Monoscope for evidence and coverage details.

Review history and settings

[claude]

Review of #429: bundle range-read bytes by index file kind.

The change is small and correct. I found no bugs, and I haven't run the tests myself.

• Attribution: BundleDirectory::read passes the path through to read_file, and StoreSource buckets the span by extension. The extension mapping matches tantivy's segment components: .idx is postings, .pos is positions, .term is the term dictionary, .fast is fast fields, and .store is the doc store.
• Fallthrough: .fieldnorm, .del and the like are not bucketed. They show up only as the remainder of range_read_bytes, as the PR description says.
• Byte counts: the counters add up the compressed bytes of the zstd block spans fetched, not the decompressed bytes. That is the right measure for object-store IO. A short note on the field doc would help whoever reads these stats later.
• Overhead: the extra cost is one extension match and one relaxed atomic add per span. That is negligible next to the S3 round trip.
• Style: it fits CLAUDE.md. The default read_file leaves the Bytes source untouched, and the diff is minimal.

Suggestions, none blocking:

1. Fieldnorm: if the remainder turns out large, add fieldnorm as its own counter. A small fieldnorm arm costs next to nothing.
2. Double counting: read_file and read are two entry points on one source, so a future caller that overrides read_file and also calls self.read could double-count. The current code is fine. A one-line comment would help.
3. Test: nothing asserts that the new counters move. A short test could open a bundle through StoreSource with a known layout, read a range from a .fast file, and check that range_read_fast_bytes equals the span and the three kind counters never sum to more than range_read_bytes. That would also cover the timefusion_stats wiring. If it's skipped because this is observability only, say so in the PR.

LGTM.

🤖 Generated with Claude Code