stephenafamo / scan / 29675347711

91%

main: 91%

Pull #13

github

feat: add WithRowSkip to skip decoding already-known rows

Preload-style queries deliver the same related row once per parent row.
A caller that shares one instance per distinct key (bob's Preload does
this since stephenafamo/bob#736) still pays to decode every duplicate:
destinations are created, scanned into and read back, then thrown away.

Add an opt-in WithRowSkip(keyColumns, seen) mapping option. Columns are
scanned in result-set order, so once the key columns are decoded, the
first remaining column asks seen whether the row is already known; if it
is, the remaining destinations are neither created nor scanned, the
validator is not consulted, and the built value carries only the key
columns — the caller promised to substitute its own copy.

Skipping degrades to normal scanning whenever it cannot apply: it needs
a TypeConverter whose destinations implement sql.Scanner (delegation),
an unresolvable key column disables it, and columns positioned before
the last key column are always decoded. Rows that are not skipped are
scanned exactly as before, so results never change.

1,000 rows x 45 string columns with 50 distinct keys (fan-in 20),
pass-through converter:

    allocs/op   92,034  ->  8,452   (-90.8%)
    B/op        5.67MB  ->  4.00MB  (-29.4%)
    sec/op      faster in every paired run (-17% .. -69%, noisy laptop)

213 of 221 new or added lines in 3 files covered. (96.38%)

742 of 817 relevant lines covered (90.82%)

43.52 hits per line