You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
rss on Linux was measuring the wrong process before this release.
getrusage's ru_maxrss survives execve, so each isolated pass inherited the parent
pytest process's high-water instead of recording its own. Every Linux rss reading was that
inherited constant — it did not vary with the benchmark at all. It now reads VmHWM from /proc/self/status, which the kernel resets for the child. macOS was never affected.
If you have saved Linux baselines containing rss, re-record them. Comparing against a
pre-0.5.0 Linux baseline will report a large, meaningless delta, and --fail-on rss:... gates
against one are not trustworthy. peak, allocated and allocations are unaffected on every
platform.
Two smaller behaviour changes in the same release: isolated passes no longer warm up, so
isolated peak is now a cold first-call number (a max_peak ceiling near its limit may newly
fail), and setup now runs before the measured pass in isolated mode.
Features
memray: add calls=N to measure buildup across invocations (#181) (7906838)
Bug Fixes
memray: don't let warmup inflate isolated rss (#180) (ff4f780)
memray: read the child's own resident high-water on Linux (#184) (b34039b)
memray: run setup before the measured pass in isolated mode (#179) (a300909)