Andrew Fribush
fig. 2.2.6 · the selector before and after its cost-model fix.
ROUTE THE AUTO SELECTOR CHOSE, PER TEST CELL · RTX 4090 · 50,000 ROWS · PLANNING FOR 100 QUERIESOne square = one test cell (dataset × threshold). Filled: screening chosen. Number: screening’s measured query speedup over dense.BEFORE THE FIXSCREENING IN 0 OF 18EXISTINGHELD OUT0.800.900.950.850.9250.9750.85×2.24×2.37×0.99×1.30×1.71×0.98×0.98×0.99×0.79×0.78×2.02×0.79×0.79×1.10×0.96×0.93×0.95×NOT RUN ON THEOLD SOURCEAFTER THE FIXSCREENING IN 18 OF 36EXISTINGHELD OUT0.800.900.950.850.9250.9751.62×2.48×2.63×2.21×2.58×2.59×0.94×1.45×1.89×1.20×1.70×2.00×0.98×0.98×0.98×0.96×0.98×0.98×0.81×2.25×3.17×0.79×2.94×3.17×0.79×0.80×2.84×0.80×2.16×3.14×0.92×0.91×0.86×0.92×0.89×0.87×ARXIVCOCOGLOVEISOTROPICCLUSTEREDDENSE OUTBold: screening measured faster, dense chosen anyway. Same source onan RTX A6000: dense in 18 of 18.No retuning before the held-out cells: screening in 10 of 18 there.RINGED CELL, OLD ARITHMETIC1 TILE TIMED,27 NOT COUNTEDONE SAMPLE TILE, DENSE2.86 msONE SAMPLE TILE, RANK-64−1.13 msSAVED PER QUERY (1 OF 28 TILES)1.74 ms× 100 QUERIES174 msCHARGED: CALIBRATION, ALREADY PAID422 msCHARGED: RANK-64 PREPARATION295 msCHARGED TO SCREENING717 ms174 < 717: DENSEPAYS BACK ONLY AFTER 413 QUERIESRINGED CELL, FIXED ARITHMETIC4 SHAPES TIMED,EACH × ITS COUNTTILE SHAPE × COUNTDENSERANK-64FULL DIAGONAL × 62.730.99FULL OFF-DIAGONAL × 152.971.28EDGE 8,192×848 × 60.430.24EDGE 848×848 × 10.100.14FULL QUERY, 28 TILES63.7226.76SAVED PER QUERY36.96 ms× 100 QUERIES3,696 msCHARGED: RANK-64 PREPARATION70 msCALIBRATION 294 ms: SUNK, NOT CHARGED3,696 > 70: RANK-64MEASURED 64.04 → 25.81 ms, 2.48×

source: gpu-threshold-join engineering-ada-2026-09-18 routes.csv, job-0010.json · portfolio-selector-2026-09-18 selector.csv, prediction-audit.json (RTX 4090) · selectorFix.data.ts

View in context: Knuth's Root of All Evil Has a Break-Even Point → · All figures

Andrew Fribush, software and computer engineer in Austin. I build consequential systems where software meets institutions: finance, healthcare, commerce, government data, and the mechanisms that govern AI agents. Taking new engagements: hello@andrewfribush.com.