Public API = the names in llm_queue.__all__, the registry document's setting
names and the llm-queue commands. Anything else may change without notice.
llm_queue.run is part of that API under this generic name: it runs a coroutine like
asyncio.run, on uvloop when the uvloop extra is installed (use_uvloop= forces it on or
off). It is kept as run, not renamed (FO-23); import it as llm_queue.run or
from llm_queue import run as run_main if the bare name clashes in your module.
0.x releases: a minor release may break the public API; every break is in CHANGELOG.md.
Message format: queue messages carry a format version v. A worker returns a message newer
than it understands to the queue instead of dropping it. Workers of two releases share a queue
only where CHANGELOG.md says so: 0.1.0b1 and 0.1.0b2 do not (drain 0.1.0b1 first; see
docs/how-to/upgrade.md).
Store schema: the key layout has a schema number; a change to it ships with a migration note.
Supported versions: Python 3.12-3.14 (unit and integration tests run on each in CI; 3.11 was dropped in 0.1.0b2); Valkey 8
and Redis 7.2+ as one primary with replicas (tested on Valkey 8.1 and Redis 7.4). Cluster mode
is not supported (one admission touches keys of several hash tags) and start() refuses it;
maxmemory-policy noeviction is recommended (a warning at start otherwise, and in strict mode
readiness fails). The client is
redis-py 6 or later (redis>=6,<9). The store must allow Lua scripts, pub/sub and blocking
list reads (BLPOP, the queue doorbell); ElastiCache for Valkey / Redis OSS does.
Dependencies: ranges with upper bounds in pyproject.toml (aiobotocore<4, redis<9,
...); a release may raise a bound, never drop one without a test on the new major.
Settings: the LLMQ_* names and the registry setting names are public API; removing or
renaming one, or changing a default that changes behaviour, is a breaking change listed in
CHANGELOG.md (under "Changed (behaviour)").