feat(router): add --disable-access-log-for-endpoints to silence probe logs

[Anai-Guo]

Summary

Closes #967.

The router container logs every liveness-/readinessProbe hit, which floods the logs. The model containers can already suppress this with --disable-access-log-for-endpoints; this PR brings the same capability to vllm_router.

What changed

• New CLI flag --disable-access-log-for-endpoints (comma-separated paths, e.g. /health,/metrics).
• A small logging.Filter (_EndpointAccessLogFilter) is attached to the uvicorn.access logger before uvicorn.run. It matches on the request path (ignoring any query string) and drops only the matching access records — error logs and --log-stats are untouched.

Why this approach

uvicorn emits each access record with positional args of (client_addr, method, path, http_version, status_code), so matching on args[2] is robust across formatters. uvicorn's dictConfig replaces handlers but preserves existing logger filters, so the suppression works in both text and json log formats. When the flag is unset, behavior is unchanged.

Usage

python -m vllm_router.app ... --disable-access-log-for-endpoints /health,/metrics

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request introduces a feature to suppress uvicorn access logs for specific endpoints (such as health probes) via a new --disable-access-log-for-endpoints command-line argument. The implementation includes a custom logging filter, _EndpointAccessLogFilter, which is registered on the uvicorn.access logger. The review feedback highlights two important issues: first, registering the filter before uvicorn.run() is ineffective because uvicorn's internal logging configuration clears existing filters, so it should instead be applied within the FastAPI lifespan context manager; second, a potential KeyError could occur if record.args is a dictionary rather than a sequence, which can be resolved by explicitly checking its type before indexing.

[gemini-code-assist]

In Python's logging.config.dictConfig (which uvicorn calls during startup), any existing filters on loggers are cleared if they are not explicitly defined in the configuration dictionary (unless incremental=True is used, which is not the case here). As a result, adding the filter to uvicorn.access before calling uvicorn.run() will cause it to be silently removed during startup, making the suppression ineffective.

To resolve this, we should store the parsed endpoints in app.state and apply the filter inside the FastAPI lifespan context manager, which executes after uvicorn has completed its logging configuration.

For example, in lifespan:

@asynccontextmanager
async def lifespan(app: FastAPI):
    # ... existing code ...
    disabled_endpoints = getattr(app.state, "disable_access_log_for_endpoints", None)
    if disabled_endpoints:
        logging.getLogger("uvicorn.access").addFilter(
            _EndpointAccessLogFilter(disabled_endpoints)
        )
    yield
    # ...

Suggested change

	# Optionally drop access logs for noisy probe endpoints. The filter is
	# attached to the ``uvicorn.access`` logger before uvicorn configures its
	# own logging; uvicorn's dictConfig replaces handlers but preserves
	# existing logger filters, so the suppression survives both text and json
	# log formats.
	disabled_endpoints = parse_comma_separated_args(
	args.disable_access_log_for_endpoints
	)
	if disabled_endpoints:
	logging.getLogger("uvicorn.access").addFilter(
	_EndpointAccessLogFilter(disabled_endpoints)
	)
	app.state.disable_access_log_for_endpoints = parse_comma_separated_args(
	args.disable_access_log_for_endpoints
	)

[gemini-code-assist]

If record.args is a dictionary (which can happen if a log message uses named placeholders), len(record.args) returns the number of keys. If the dictionary has 3 or more keys, record.args[2] will raise a KeyError because 2 is not a valid key. To prevent potential logging crashes, we should explicitly verify that record.args is a sequence (like a tuple or list) before indexing it.

Suggested change

	def filter(self, record: logging.LogRecord) -> bool:
	if self._endpoints and record.args and len(record.args) >= 3:
	path = str(record.args[2]).split("?", 1)[0]
	if path in self._endpoints:
	return False
	return True
	def filter(self, record: logging.LogRecord) -> bool:
	if (
	self._endpoints
	and isinstance(record.args, (tuple, list))
	and len(record.args) >= 3
	):
	path = str(record.args[2]).split("?", 1)[0]
	if path in self._endpoints:
	return False
	return True