Module stamps 

Receive-time stamps for a range of transcript lines, from the output.tsidx sidecar — without loading the sidecar whole.

The sidecar holds one record per append, {"off":<byte offset>,"ms":<unix ms>}\n, in append order, so off rises through the file. A line’s stamp is the newest record at or before the line’s own byte offset (0 = unknown).

blockfile:read_range used to read the whole sidecar and parse every record to stamp ~200 lines: 38 MB and ~1.1M records on a long-lived agent, about 0.6 s per read. Instead, a binary search over the sidecar’s bytes finds the record the first line needs, and only a window around the requested range is read and parsed.

Records a little out of order (two srv instances mirroring into one global zone, older builds stamping outside the append’s transaction) are tolerated: the window reaches MARGIN bytes past both ends of the range and its records are sorted. A record further out of place than that can be missed, so a line may get an earlier stamp — stamps are best-effort display times, and a miss that far out would need thousands of records out of order.

Structs

Record 🔒

Enums

Probe 🔒

Constants

CHUNK 🔒

Chunk size for the window read.

MARGIN 🔒

Bytes read before the range’s first record and past its last, for out-of-order records (~2,000 records at ~30 bytes each).

PROBE 🔒

Bytes read at each binary-search probe; a record is ~30 bytes.

SEARCH_STOP 🔒

Once the search interval is this small, the window just covers it.

Functions

first_record_after 🔒

The first parsable record that starts after pos (at a line boundary strictly past it) and before hi.

parse 🔒

stamps_for 🔒

Stamps for line_offsets (ascending byte offsets in output), read from a sidecar of size bytes through read(offset, len) — None from read means those bytes aren’t all stored (a writer mid-append), and so no stamps. Ok(None): no sidecar records near the range at all.