Module resource_usage 

What a runner’s process tree uses right now: resident memory and CPU time summed over a root process and every descendant (coordinator stage 4, job weight statistics).

One call is one cheap look, taken every few seconds by the daemon off its reconcile loop:

• Linux reads /proc/<pid>/stat for every process (parent, CPU ticks, resident pages). A native runner does not get a cgroup of its own, so the tree is the unit.
• macOS lists every PID (proc_listallpids), reads each parent (PROC_PIDT_SHORTBSDINFO, which any account may read, so a sudo step keeps its place in the tree) and, for the tree, PROC_PIDTASKINFO (resident size, user and system time in Mach ticks).
• Windows takes one Toolhelp snapshot for the parent links and, for the tree, GetProcessMemoryInfo (working set) and GetProcessTimes. A native runner is not started inside a Job Object, so there is no job accounting to read. Toolhelp keeps a dead parent’s PID, which Windows soon hands to another process, so a child created before its supposed parent is not counted as its child.

The CPU total of a tree drops when a descendant exits (except on Linux, which folds a reaped child into its parent’s cutime/cstime); the caller adds up the increases between looks, so at most a look’s worth of an exiting process’s time is lost. A process this account may not inspect counts as using nothing. Work that leaves the tree is not seen: containers run by dockerd, and daemons that detach (a Gradle daemon, a nohup server) and are re-parented to init.

Structs

TreeUsage

One look at a process tree.

Functions

sample_tree

The tree under root, root included, or None when root is gone.

sample_trees

The trees under each of roots, from one look at the process table: a root that is gone has no entry.