[shikhar]

cachey_bucket_latency_* reports every latency model (first_byte, body, request) at three quantiles: up to 48 series per bucket per instance. Dashboards read only the request model's median and p99. The first-byte and body models matter for what they set, the early hedge delay, but that delay is the smallest of p99, 1.25 × p95, and 5 × the median, so it could not be read off any one series.

Change

• Bucket latency, request model only: cachey_bucket_latency_seconds, _model_bytes, _predictions_total, and _predictions_exceeded_total now report only the request model, and latency only at the median and p99. Metric names are unchanged and phase="request" stays, so existing queries keep matching.
• New cachey_bucket_hedge_delay_seconds{bucket, phase, bytes}: the delay before an early hedge fires, at the reported read sizes up to the largest the model retains. phase is first_byte (timed by requested bytes) or body (timed by body bytes), matching cachey_hedges_total. It is the value the downloader uses.
• Series count: up to 20 per bucket, down from 48: 12 for the request model and 8 for hedge delays.
• BucketMetrics: carries hedge delays instead of the first-byte and body estimates.
• cachey_fetch_request_entries: buckets at 1, 2, and 4 instead of 1 to 32. That's 6 series per kind instead of 9.
• cachey_fetch_request_bytes: buckets from 4 KiB to 16 MiB, four times apart, as the latency model spaces its bands. That's 10 series per kind instead of 9. The lowest bound was 512 KiB, so every small exact-range read fell in the first bucket.
• Test removed: the integration test that only checked a ranged GET counts status="success". Other integration tests cover ranged reads.

docs/replica-reads.md documents the new gauge and the request-only latency series.

Validation

• 204 tests, cargo +nightly fmt, and Clippy with -D warnings pass.
• A throwaway check, not committed, fit a model to 4 KiB, 64 KiB, and 1 MiB reads. Its reported delays matched hedge_delay at those sizes, and 16 MiB was omitted rather than extrapolated.

🤖 Generated with Claude Code

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge.

Summary

This PR reduces exported bucket latency series and changes fetch-size histogram bounds.

• Bucket latency metrics show request times at the median and p99.
• Fetch histograms group bytes and entries into new buckets.
• The ranged-GET metrics check is no longer in the integration suite.

Acknowledgment: shikhar explicitly described the test removal and cited other ranged-read integration tests as the reason.

_{Reviews (1) · Last reviewed commit: "feat(service): bucket fetch sizes from 4..." · Reviewed by Greptile}