v2.0.1

Performance

• _to_pyformat() cached with functools.lru_cache(4096) — the
  $N → %s placeholder rewrite is on every PG query's hot path.
  Real apps reuse the same SQL strings billions of times across
  requests; caching converts the per-call O(len(sql)) state-machine
  into a dict lookup.
• PG __in lookup uses = ANY(%s) instead of
  IN (?, ?, ...). One prepared-statement shape regardless of the
  list length, so PostgreSQL's plan cache hits across calls with
  different len(ids). SQLite stays on the classic IN syntax.
• M2M add() / aadd() batched into 2 queries — previously
  for obj in objs: SELECT 1; INSERT VALUES (..) issued 2N
  round-trips. Now one SELECT ... WHERE tgt IN (...) to find
  existing links + one multi-row INSERT for the missing ones.
  Adding 1000 tags drops from ~2000 queries to 2.
• M2M remove() / aremove() batched into 1 query —
  DELETE ... WHERE tgt IN (...) instead of N per-object DELETEs.
• Async prefetch_related parallelized via asyncio.gather —
  previously each prefetched relation awaited sequentially, so
  .prefetch_related("author", "category", "tags") cost 3× the
  latency of one. Now they fire concurrently; total wait collapses
  to the slowest single sub-query.
• bulk_create() / abulk_create() field list hoisted out of the
  batch loop — small constant-factor win (~5%) when the call has
  many batches, and clearer code.

Added

• PostgreSQL pool PREPARE_THRESHOLD setting —
  DATABASES["default"]["PREPARE_THRESHOLD"] is forwarded to
  psycopg's connection kwargs. Set 0 for "always prepare"
  on workloads dominated by repeated SELECT/UPDATE shapes; leave
  unset to keep psycopg's default of 5. Both sync and async wrappers
  honour it.