MemoryTelemetry::Poke() runs after every main-thread task (from XPCJSContext::AfterProcessTask, https://searchfox.org/firefox-main/rev/6c74efe2fcddf84b6f320959064a66946b4a1759/js/xpconnect/src/XPCJSContext.cpp#1542). Ten tasks within one second count as "active" and arm a one-shot timer of kTelemetryIntervalS (60 s, https://searchfox.org/firefox-main/rev/6c74efe2fcddf84b6f320959064a66946b4a1759/xpcom/base/MemoryTelemetry.cpp#44). A content process trivially produces ten tasks per second even while idle in the background, so in practice every live process gathers once a minute for as long as it exists, and the parent additionally reads the resident-unique size of every child in GatherTotalMemory on the same cadence.

The gather is not cheap, because the resident-unique measurement walks the whole page table of the process. Bug 2069230 switched Linux to smaps_rollup, which removed the text formatting and parsing but not the walk. Weighted medians of memory.collection_time from Nightly telemetry, three days before versus three days after that landing:

The cost was large enough to show up as a 69% improvement in background-resource cpuTime-tab on android-hw-a55 (perf alert 52774): before bug 2069230 roughly 70% of the CPU time of an idle background tab in that ten-minute test was spent gathering memory telemetry. Fenix Nightly clients record about 200 tab-process samples per foreground hour.

Bug 2034081 wants to add per-process thread statistics to the same gather.

Note on channels: on Android this only affects Nightly (and the Nightly-flavour CI builds), because Fenix release does not run unified telemetry and Telemetry::CanRecordReleaseData() stays false there, so MemoryTelemetry never gathers (2 of 1.18 M Fenix release clients had a memory.unique sample on 2026-09-13). Desktop is different: unified telemetry keeps recording on for every channel, and on 2026-09-13 90-99% of desktop release clients gathered, about 3200 (Windows) to 5000 (Linux) times per client-day. At the current Linux median of 37 ms per gather that is roughly three minutes of CPU per Linux release user per day.