mirror of
https://github.com/stablyai/orca.git
synced 2026-11-15 18:37:24 +00:00
The loader no longer decodes the table into one string, a string per line and an array per range before packing it: it scans the (gunzipped) bytes a line at a time, decodes only a line's first three fields, and writes typed arrays sized from the line count, sorting an index permutation only when the file is out of order. Behaviour is unchanged (same fields, trimming, quotes, codes, stable order on equal starts; 600k random lookups agree with the previous loader on DB-IP 2026-11). Measured on that table in a linux/amd64 container: cgroup peak 312 -> 124 MB, RSS after load 260 -> 151 MB.