Module sysinfo 

Memory detection for auto-budget computation.

Production target is Linux (Fargate/ECS). On Linux we walk the cgroup hierarchy to find the container memory limit, then fall back to host RAM reported by /proc/meminfo. On macOS we read host RAM via sysctl; cgroup support requires unsafe FFI so it is deferred.

cgroup-v2 detection on ECS Managed Instances (Bottlerocket)

ECS Managed Instances launch containers without --cgroupns=private, so /sys/fs/cgroup/memory.max resolves to the unified-hierarchy root, which reads "max" (no limit). The actual container memory limit is set at a deeper path whose last component is recorded in /proc/self/cgroup (unified-hierarchy format: a single line 0::<path>, e.g. 0::/ecs.slice/ecs-…-task.scope/<id>).

cgroup_memory() reads /proc/self/cgroup, extracts that path, then reads memory.max at each ancestor (deepest first) until it finds a numeric limit or exhausts the tree. Falls through to host_ram only when the entire walk yields "max" (truly unconstrained host).

RSS tracking (read_process_rss_bytes) is Linux-only (parses /proc/self/statm). On macOS it returns None; the auto-budget logic treats None as “cannot measure model footprint” and uses conservative defaults.

Structs

MemoryReading

A memory reading with its provenance.

Enums

MemorySource

Where the available-memory reading came from.

Functions

cgroup_memory 🔒

cgroup_v2_walk 🔒

Reads the cgroup v2 memory limit by walking ancestors of the container’s cgroup path.

detect_available_memory 🔒

Detects available memory for the process.

env_override 🔒

host_ram 🔒

Linux: parse MemAvailable from /proc/meminfo (kB → bytes). macOS: read total host RAM via sysctl hw.memsize.

linux_meminfo_available 🔒

linux_rss 🔒

page_size_bytes 🔒

read_process_rss_bytes 🔒

Returns the current process’s RSS (Resident Set Size) in bytes, or None if measurement is not supported on this platform.