DBIx-Fast-0.21

Changes for version 0.21 - 2026-10-06

Connection registry: configure(name => {...}) + connect(name), one instance per name per process (rebuilt after fork)
stream(): query() without loading the result set into memory (MariaDB/MySQL use_result; 200k rows: +45 MB -> +0 MB)
cached(): tag invalidation shared across processes with an external backend (tag versions in the backend); DBIx::Fast::Cache::Redis backend over Redis::Fast. With an external backend, invalidate() now returns 1 instead of a per-process key count
after_write => sub {...}: hook called after every CRUD write (audit)
DBIx::Fast::Test::Mock: unit tests without a server (canned answers, real SQL of the chosen dialect, recorded statements)
on_slow_query => { ms, cb }: report slow statements
metrics => 1, metrics(), metrics_text(): statement counters, duration histogram and errors by class in Prometheus format
configure_file(): registry connections from a JSON file
Writes 32-42% faster than 0.20 (insert/update/delete/upsert) and all() 16-19% faster on SQLite; 5-11% on MariaDB: no per-write closures, fewer method calls, validated-identifier memo, rows built from the driver's C fetchall
Pg: insert/insert_ignore return the id with INSERT ... RETURNING (one round trip; last_insert_id cost another: insert ~2x faster on Pg)
val()/hash()/all() 31/24/9% faster than the first 0.21 build on SQLite: fast path for cached statements, and Connector::dbh's common case in a method touching 3 fields instead of 11 (Object::Pad binds every field a method mentions): 428 -> 282 ns per query
Bounded caches (statements, CRUD SQL, identifiers) shed a quarter of their entries when full instead of emptying (no burst of misses)
json_columns: write column sets memoized per table; reads decode only the configured columns the result actually has
DBD::SQLite, Test::More, Test::Exception are test-only prerequisites (TEST_REQUIRES) - no longer installed for a runtime install