textBoundBytes

Byte length at which a text bound is truncated.

A path whose values are whole documents would otherwise put two of them in every sketch. The truncation always widens: a minimum is cut to a prefix, which cannot be larger than the value it came from, and a maximum is cut and then incremented, which cannot be smaller. So a truncated bound stays a correct bound.

This dial does not decide what a query skips, and an earlier version of this comment said it did. A sketch bound is descriptive: it is rendered by InferredSchema.render, readable through InferredField.bounds, and written to the .cat sidecar. Nothing in rabosh-query reads it. Skipping is decided by a shredded column's bounds, which are built from the same ValueBoundsBuilder — that is the sharing this module exists to enforce — but truncated at IndexOptions.columnTextBoundBytes, a separate dial in rabosh-index that also happens to default to 64. Widen that one to buy pruning; widen this one to see more of a value when you look at the model.

A corpus of long shared-prefix strings is where either default stops earning its keep, and it is worth naming because it is a common shape rather than a pathological one. Protobuf-JSON @type values begin type.googleapis.com/, which is 20 of the default 64 bytes: type.googleapis.com/com.example.game.player.v1.PlayerDTO is 56 and still discriminates, one package deeper does not, and at that point every bound is the same prefix. Here that costs legibility; on a column it costs pruning, and :rabosh-bench:runTextBoundCost is the sweep that prices it. Nothing about either is a correctness question — truncation widens — so both are tuning.