OOM kills

Processes killed by the kernel out-of-memory killer, from /proc/vmstat's oom_kill. A cgroup reaching its own memory limit counts here too, not just a host that ran out.

2026-08-31 21:13:23 UTC → 2026-09-01 03:13:23 UTC (6h 0m wide, 5m 0s rows) · last poll 3m 20s ago

unit: kills per second · min/avg/med/max/σ/p5/p25/p75/p95 over the window shown · σ and the percentiles show '-' where the window held too few samples for them to mean anything, the cell naming how few · Δ vs the preceding equal window · a paired column sorts by either half
field — the plugin's label for the field, or its name where it gave none; the swatch is the colour the graph drew it with	cur — the newest consolidated value in the window shown	min — lowest value over the window shown — where the window is served by a coarser archive this is the true minimum, read from the RRD's own MIN archive rather than taken from the averaged rows; a field with a cdef stays on those averaged rows, since the minimum of a difference is not the difference of the minima	avg — mean of the consolidated values over the window shown	med — median of the consolidated values over the window shown — always an actual sample (the upper middle on even counts, never an interpolated value); beside avg so skew is visible at a glance	max — highest value over the window shown — where the window is served by a coarser archive this is the true maximum, read from the RRD's own MAX archive rather than taken from the averaged rows, so a spike inside one consolidated bin still shows; a field with a cdef stays on those averaged rows, since the maximum of a difference is not the difference of the maxima	σ — population standard deviation of the consolidated values over the window shown — an approximation over consolidated rows on long windows, unlike min and max, which read the RRD's own MIN/MAX archives there (except on a field with a cdef)	p5…p95 — 5th and 95th percentiles (linear interpolation) of the consolidated values over the window shown — the robust floor and ceiling beside min's and max's literal ones, for the fields where one unlucky bin is not the story; sort by either half. An approximation over consolidated rows on long windows, unlike min and max, which read the RRD's own MIN/MAX archives there (except on a field with a cdef)	p25…p75 — 25th and 75th percentiles of the consolidated values over the window shown — the box around med: half the window's samples lie between them, so the spread they show is one no single excursion can move, where σ beside them is a squared distance that skew drags around	Δ — this window's average against the preceding equal window's, as a signed percentage; consolidation may differ between the two windows on fine-rung boundaries, and a cell whose two windows came back at different resolutions says so on hover; '-' when the previous window offers no baseline — hover for why	trend (per hour) — least-squares slope of the drawn values over the window shown, per hour — hover a cell for how well its line fits the rows behind it	threshold in — projected time until the stored value crosses a warning or critical bound, from a least-squares fit of the stored series; each cell is earliest plausible / point estimate / confidently expected using a ±2 standard-error prediction interval that widens further ahead; the last may be — within the four-window forecast horizon; '-' has its reason on hover	warn/crit — the values that breach this field to WARNING and to CRITICAL, in that order; hover either half for the range as configured	limits — the plugin's declared min/max: values outside this range are stored as UNKNOWN rather than clipped (rrdcreate's DS min/max); hover a cell for the exact bounds	state·since — the verdict at the last poll, and how long the field has been in it — hover the word for what it means; a steady OK has no transition on record and shows none	info — the plugin's description of the field, and its own note about a breach
OOM kills ▤	0.00	0.00	0.00	0.00	0.00	0.00	- — 4 known samples in this window, fewer than the 21 p5 needs: below that neither the 5th nor the 95th percentile rank sits a whole sample inside the extremes, so both collapse onto min and max…- — 4 known samples in this window, fewer than the 21 p95 needs: below that neither the 5th nor the 95th percentile rank sits a whole sample inside the extremes, so both collapse onto min and max	- — 4 known samples in this window, fewer than the 5 p25 needs: below that neither quartile's rank sits a whole sample inside the extremes, so both collapse onto min and max…- — 4 known samples in this window, fewer than the 5 p75 needs: below that neither quartile's rank sits a whole sample inside the extremes, so both collapse onto min and max	-	0.00	- — the trend heads away from every configured bound	not 0/-	≥ 0	OK·-	Processes killed by the out-of-memory killer, including per-cgroup kills from a unit or container hitting its own memory limit. Warns on any occurrence: the host survived to report it, but something did not.

dark mode inverts graph colours; field-table swatches, where shown, are inverted too so they match the curves