amazon-aurora / babelfish_extensions / 25571405065

77%

main: 77%

push

github

[OSS-ONLY] Add persistent ANTLR parse tree cache for routines (#4547)

Cache serialized parse trees in sys.babelfish_function_ext to eliminate
redundant ANTLR parsing on first execution in new sessions. The PLtsql
function hash table is session-scoped, so without caching every new
session re-parses large routines from scratch. This reduces first-
execution time from ~2s to ~15ms for routines with ~1300 lines.

Serialized trees use a nodeToString/stringToNode framework modeled after
PostgreSQL's node serialization, with a Perl code generator producing
outfuncs/readfuncs/equalfuncs from annotated PLtsql headers. Extension-
owned NodeTags are offset from 10000 to avoid collision with PG's enum.
Cache is written at CREATE/ALTER time, restored on first EXEC in a new
session, and validated against BABELFISH_VERSION_STR to reject stale
entries after upgrades. Cache writes are skipped on read-only instances.

Caching behavior is controlled at three levels: a session-level GUC
(babelfishpg_tsql.enable_antlr_parse_cache, default true) provides the
global toggle, a per-function setting toggled via
sys.enable_antlr_parse_cache(routine_oid, true/false/NULL) allows
function owners to force-enable, force-disable, or follow the session
GUC, and a superuser-only debug GUC (validate_antlr_parse_cache)
enables roundtrip verification by comparing cached trees against fresh
ANTLR output.

Limitations: trigger functions and event triggers are not cached in this
initial implementation due to differences in datum array layout between
CREATE and EXEC time. These will be addressed in a follow-up.

Task: BABEL-6037

Author: Manisha Deshpande mmdeshp@amazon.com
(cherry picked from commit 56b641e7c)

582 of 909 new or added lines in 16 files covered. (64.03%)

53876 of 69915 relevant lines covered (77.06%)

643498.94 hits per line