When local is enough — and when it isn't

Decision lab: five limits end the local shape and only three of them are a move. The store is 24 records, 9 namespaces, 17 fields, 408 field slots — 36,864 bytes of vectors, 9 pages, 0.024% of a 100,000-vector brute-force ceiling that costs 146.48 MiB and 38,400,000 multiply-adds a query. The ceiling is a date: 49,988 sessions away at 2 records a session, 24,994 at 4, 16,663 at 6, and 417 at a nightly 240. A corpus problem asks for an index and a writer problem for a lock, so both stay on one file: 80 whole-file rewrites from two writers lose 40, and one lock loses none. What moves is a second machine, one namespace on two machines, and the 117 of 168 hours a week the laptop is off. Whatever moves, the artefact is the schema: records only keeps 48 of 408 slots (11.8%), vectors only 72, and the contract 384 (94.1%) with the vector rebuilt — a round trip that compares 408 slots and finds 0 differing. Of six scenarios, four move and two stay, and the vendor benchmark is one of the two.

This Memory Engineering lesson is part of a paid plan. Sign in with a plan that includes this course to read it.