Memory numa stats #2621
Memory numa stats #2621
Conversation
Signed-off-by: Katarzyna Kujawa <katarzyna.kujawa@intel.com>
metrics/prometheus.go
Outdated
| values := make(metricValues, 0) | ||
|
|
||
| values = append(values, getNumaStatsPerNode(s.Memory.ContainerData.NumaStats.Total, | ||
| []string{"total", "container"}, s.Timestamp)...) |
There was a problem hiding this comment.
The general rule of thumb for when to use a label vs when to add a new metric is that the sum of a metric across all dimensions should be meaningful. "total" isn't a great dimension to have, as we would expect the sum of dimensions to be the "total". So we can either calculate the "other" portion, or make ...pages_total a separate metric.
There was a problem hiding this comment.
In documentation it is written that "total" count is sum of file + anon + unevictable so I'll remove metrics with "type"="total".
metrics/prometheus.go
Outdated
| []string{"unevictable", "container"}, s.Timestamp)...) | ||
|
|
||
| values = append(values, getNumaStatsPerNode(s.Memory.HierarchicalData.NumaStats.Total, | ||
| []string{"total", "hierarchy"}, s.Timestamp)...) |
There was a problem hiding this comment.
I suspect hierarchy vs container may need to be separate metrics by the same logic above.
There was a problem hiding this comment.
I followed the pattern which was used for "container_memory_failures_total metric", see this. There is "scope" label with values "container" or "hierarchy".
| name: "container_memory_numa_pages", | ||
| help: "Memory usage per numa node", | ||
| valueType: prometheus.GaugeValue, | ||
| extraLabels: []string{"type", "scope", "node"}, |
There was a problem hiding this comment.
I think we only add two labels to metrics below?
There was a problem hiding this comment.
I think this comment is still relevant
| if includedMetrics.Has(container.MemoryNumaMetrics) { | ||
| c.containerMetrics = append(c.containerMetrics, []containerMetric{ | ||
| { | ||
| name: "container_memory_numa_pages", |
There was a problem hiding this comment.
is this measured in #pages, or bytes? The help text should specify the units, and the suffix of the metric should be _
Improve help text for prometheus metric Signed-off-by: Katarzyna Kujawa <katarzyna.kujawa@intel.com>
|
@dashpole Could you take a look if new names of metrics are better? |
Signed-off-by: Katarzyna Kujawa katarzyna.kujawa@intel.com
This pull request introduces information from memory.numa_stat as Prometheus metrics.