Highlights
Denial of Service: 20 prior fixes. Scrutinize any change in this area.
omlx/scheduler.py: most-fixed (8 issues). Treat as high-risk during review.
5 high-severity fixes in this history; regressions here are high-impact.
Recurring patterns
The bug types that recur here, drawn from past fixes, not open vulnerabilities.
Denial of Service: The scheduler is the primary point of failure for memory exhaustion. Out-of-memory errors, failure to release paged caches on rejected requests, and lack of synchronizations during cache clearing have repeatedly triggered process crashes (SIGABRT) or kernel panics. Developers must ensure all allocation errors cleanly release reserved blocks.
Denial of Service: Disconnected system components (such as unassigned monitors or nested config mismatches) have historically allowed huge prefill requests to bypass memory guards entirely, causing hardware-level Apple IOGPUFamily kernel panics. Thorough integration tests are needed to verify end-to-end enforcement.
Denial of Service: Race conditions during engine management can cause premature unloading of active models or build up un-tracked memory pressure. Developers must ensure that engines are only evicted after active client leases are fully drained and that backend buffers are fully freed.