Until this month the biggest standing test collection was 100,000 vectors. The first real end-to-end run at a million found two P1 bugs in the first day. Both invisible below a million, by construction.
The first was in the seeder -- the part that queues index experiments for the optimizer to try. It deduplicated only terminal experiments and topped the queue back up every 60 seconds. On a small collection the queue drains faster than it refills, so you never notice. On a million vectors the drain is slow, the refill outran it, and the queue grew to 75,483 rows. Convergence became unreachable -- the optimizer can't finish a grid that keeps regrowing. Fix was a queue-empty gate: drain before you refill. The next run held flat at 436 rows the whole way.
The second was nastier. Under sustained load, control-plane SQLite contention at a census boundary stalled data-plane queries, and an eight-hour run died at minute 505. Also scale-only -- the contention needs enough real traffic to collide with the bookkeeping.
Neither was a mystery once found. Neither would EVER have shown up at 100k. The standing test collection is a million now.