From Snakes to Crabs

-- ONE query per request. Conditions for EVERY flag,
-- evaluated inline by Postgres:
SELECT person.id, person.properties,
  (person.properties->>'email' LIKE '%@acme.com' AND
   (person.properties->>'plan')::text = 'enterprise') AS flag_a,
  (cohort.id = ANY($2)) AS flag_b,
  -- ... one expression per flag, for every flag
FROM person
  LEFT JOIN cohortpeople ON cohortpeople.person_id = person.id
  LEFT JOIN cohort ON cohort.id = cohortpeople.cohort_id
WHERE person.team_id = $1 AND person.id = $3;

# Postgres caps out in the low hundreds, so thousands of
# workers funnel through PgBouncer. Our deadlines live there too:
query_timeout      = 15
query_wait_timeout = 10

# in the Django request handler:
# …no real per-request deadline we could reach.

// sqlx owns the pool — no PgBouncer in the path.
let pool = PgPoolOptions::new()
    .max_connections(20)
    .acquire_timeout(Duration::from_millis(50))
    .connect(&url).await?;

// And the deadline lives in our code, per request:
let row = timeout(Duration::from_millis(100), q.fetch_one(&pool)).await?;

// Fetch raw properties once, then fan the matching
// out across cores — the thing the GIL never allowed.
let results: Vec<FlagMatch> = flags
    .par_iter()                       // rayon
    .map(|flag| flag.evaluate(&props))
    .collect();