What text search forbids

You cannot simultaneously have exact global top-k ordering, sub-second visibility, an arbitrary scoring function, and exact faceting at low cost. Each pair is fine; the set is not. Specifically:

order costs positions proportional to tokens;
a non-decomposable score costs all skipping;
facets cost a forward copy and cancel top-k pruning;
freshness costs segment count, and therefore per-query fixed work;
global score statistics cost a coordination round.

The escapes and their prices

Escape	What it buys	Price
Multi-stage retrieval	makes an arbitrary scoring function affordable by confining it to a few hundred candidates	a hard recall ceiling set by the cheap stage, and a tuning surface with no closed form
Precomputation (n-grams, materialized facet aggregates, multiple analyzed fields)	converts read work into write work and space	vocabulary or index-size blow-up
Approximation of facet counts	buys facet aggregation back	two-sided error, licensed only because facet counts are advisory (the candidate-generation approximations in search are one-sided by necessity)
Local IDF	buys away a coordination round	silent, unmeasurable score inconsistency

Case, resolvedConstructed case · Amara at Haatly

One extra word, 200 times slower

The new search cluster took 1.9 s for "the red kurta", pages with counts were slow, new listings were invisible for a while, and searching Apple found apples. Where it started: Text search →

"The red kurta" at 1.9 s. "The" has a gigantic postings list with a useless score bound (nearly every document contains it, and its score contribution is tiny and flat). Amara adds "the" to the stop list for ranking purposes and relies on block-max skipping for the rest. The query returns to ~10 ms. (A stop word is an index-time decision, so she accepts that exact-phrase queries like "the who" need an unstemmed, unstopped second field.)

Slow pages with counts. She stops counting everything. Facets are computed on a sample of the match set and shown as "about", with exact counts only for the low-cardinality category field.

Invisible listings. She keeps the one-second refresh for the catalogue but merges segments more aggressively, so there are never more than a handful of segments per shard. At 100 ms refresh the segments multiplied and query time went up for everyone, exactly as the per-segment fixed cost predicts.

"Apple". The default stemmed field loses case. She adds an exact, case-preserving field and boosts matches on it. Cost: a larger index and a rebuild.

She also sees a pattern: an index is wrong in one direction, a fixed cost is paid by someone, and the cheap stage limits the expensive one.