v1.17.1

• benchmark/db_users: a Users API benchmark sub-project — an APIRoot with
  reflection-wired modules and Role/Address/User entities, measured end
  to end against memory, SQLite or PostgreSQL (--docker or an existing
  server), with latency percentiles and a results history (HISTORY.md).

• Performance, found and measured with benchmark/db_users:

  • Entity tracking: a loaded entity's referenced entities (and lists of
    them) are snapshotted by ID, not by a reflective deep copy. Change
    detection already compared them by ID. SQLite: reads +22%, lists +40%,
    updates +30%.
  • PostgreSQL: the reads of a transaction that hasn't written yet run
    without BEGIN/COMMIT — a select of an entity with references was
    BEGIN, its queries and COMMIT. Its first write opens the transaction.
    DBSQLAdapter.readsOutsideTransaction, on by default for PostgreSQL
    (READ COMMITTED), opt-out with readsOutsideTransaction: false.
    Single-entity reads 2x faster.
  • DBSQLMemoryAdapter: relationship selects and inserts go through the
    relationship indexes; a delete by ID takes its row by key and resolves
    only it; a relationship delete is narrowed through the index. Reads up
    to 3x, writes up to 7x faster.

• Fixes:

  • LoggerHandler.root threw a LateInitializationError when it was the
    first logging access.
  • DBSQLMemoryAdapter: after a rollback, a relationship restored in its
    table was missing from the relationship index (an index entry is a Set
    changed in place, which MapHistory can't roll back), so queries through
    the relationship missed it. Its indexes versions snapshot was also never
    refreshed.

• Dependency updates:

  • reflection_factory: ^2.10.1 (faster constructor lookup when decoding
    entities: +10–15% on reads)
  • map_history: ^1.0.7 (consolidate visits only the changed keys: the
    memory adapter's commits no longer scale with the table size)