feat(schema): make the redis graph scan skippable, and fix skip_redis

[YaphetKG]

Problem

Building the schema scans the redis graph for node types and a summary. Measured against a 76M-edge FalkorDB graph:

	measured
full schema build	833 s
of which the summary query MATCH (c) RETURN DISTINCT labels(c) as types, count(c) as count	99 s
node types resolved	41

Three multipliers on top of that:

• SchemaFactory._cached is a class attribute, so the build is per worker process.
• The refresh thread repeats it every update_interval, hardcoded to 20*60. An 833 s build on a 1200 s sleep means each worker spends ~40% of its life rebuilding the schema, indefinitely.
• A worker's first request pays it twice concurrently — SchemaFactory.__init__ builds synchronously, while the update_cache_loop thread it just started immediately builds again.

Observed effect: raising gunicorn from 1 to 16 workers made a client crawl make zero progress for 30+ minutes, with the tranql pod pinned at its 2-core limit and redis idle at 7m.

None of that scan is used by redis: queries

SelectStatement.execute reads only:

all_schemas = interpreter.schema.config['schema']
redis_key = [k for k in all_schemas if 'redis' in all_schemas[k] and self.service.startswith(k + ':')]
redis_connection_details = all_schemas[redis_key]['redis_connection_params']

That's schema.config, loaded from schema.yaml. It then hands the query graph to PLATER's GraphInterface.answer_trapi_question. The scanned type map lives in schema.schema, which this path never reads — generate_questions uses the parsed query's concept.type_name/concept.curies, and plan()/execute_plan() is the /schema branch.

skip_redis already existed, but was unreachable and broken

• Unreachable: TranQL.__init__ constructed SchemaFactory(...) without it, so it was always the default False.
• Broken: the scan is also what sets metadata['schema'] (line 309), and schema_data = metadata['schema'] is read unconditionally later in the same loop (line 322). So skip_redis=True raised KeyError('schema') for any redis-backed entry — it can only ever have run against configs with no redis schema. My test caught this; skipping now leaves an empty type map rather than an absent key.

Change

• Wire SKIP_REDIS_SCHEMA through TranQL.__init__ → SchemaFactory → Schema, and make the skip path coherent.
• Expose SCHEMA_UPDATE_INTERVAL, because 20 minutes is only safe relative to how long the scan takes.
• Add util.as_bool. Config.__getitem__ returns environment overrides verbatim, so SKIP_REDIS_SCHEMA=false arrives as the string "false" — and bool("false") is True. Without coercion the off switch would mean on. It raises on anything uninterpretable rather than silently picking a branch.

Default stays false, so current behaviour is preserved. Deployments whose queries all name a redis: source can turn it on.

Caveat

TranQL's own /search endpoint infers its default indexes from the scanned schema (api.py: schema.schema[redis_schema_name]["schema"].keys()). Enabling SKIP_REDIS_SCHEMA leaves that empty, so it should stay off for any deployment using /search. The query endpoint is unaffected.

The other flags read via Config.get (ASYNCHRONOUS_REQUESTS, NAME_BASED_MERGING, …) have the same latent string-boolean issue if set from the environment. Not touched here to keep the diff reviewable.

Tests

tests/test_skip_redis_schema.py, 22 tests, no redis or graph needed:

• as_bool across "false"/"False"/"0"/"no"/"off"/"" and the truthy forms, real bools, None + default, and that garbage raises
• skip_redis=True does not construct RedisAdapter, while config['schema'][...]['redis_connection_params'] still loads — that's what the query path actually reads
• skip_redis=False still constructs it, so this fix cannot silently disable the scan for everyone

tests/test_tranql.py is 8 failed / 41 passed both before and after, verified by running the same suite against a pristine origin/develop tree — the same eight test names fail either way, so they are pre-existing.