Support label_replace() on range vectors?

[juliusv]

Sometimes you might want to do something like rate({__name__=~".*_total"}[1m]), but since this would lose the metric name, you might want to copy the metric name into a normal label as the very first step. Currently this is not possible with label_replace() for any functions that expect a range vector, as label_replace() only supports relabeling an instant vector.

While you can do:

label_replace({__name__=~".*_total"}, "metricname", "$1", "__name__", "(.*)")

...the following is currently illegal:

label_replace({__name__=~".*_total"}[1m], "metricname", "$1", "__name__", "(.*)")

But you would need this if you wanted to pass the result into a function that requires a range vector, as you can't specify a range selector ([1m]) after arbitrary expressions, only basic series selectors.

It's unclear to me though whether we want functions to take varying argument types, or how we would solve that if not.

[brian-brazil]

I'm against having a function of this nature.

It's expected that you know your metric names, and all uses cases we've seen for this so far were where the user had encoded a label value into the metric name. I don't think we should be adding features to work around bad data modelling - particularly as we already have metric relabelling.

I'm against functions' return types and semantics varying depending on the type of input arguments, that's just going to cause confusion.

Finally we currently have no functions that return range vectors. I'd like to keep things that way, particularly as I expect that'll mess up optimisations that LTSes will be doing.

[juliusv]

Ok, agreed for now - let's see if a more urgent need for this comes up ever again. Closing for now.

[fortman]

So the use case still exists where you want to change the value of a label, having nothing to do with a metric name. Our use case:

We are using cadvisor to pull in docker statistics. We are showing the image name in a grafana dashboard. The problem is our docker image names are fully qualified so they are on the long side and hard to read in grafana. The naming convention is outside my control, so when I found the label_replace function I got really excited. The following worked:

label_replace(container_cpu_user_seconds_total_counter{image!=""}, "image", "$1", "image", "./(.)")

However our grafana dashboard requires a ranged vector(because we are doing an irate), and the the following does not work

label_replace(container_cpu_user_seconds_total_counter{image!=""}[1m], "image", "$1", "image", "./(.)")
Error executing query: parse error at char 106: expected type instant vector in call to function "label_replace", got range vector

Was excited, but now I get to live with unreadable grafana dashboards and management on my .... well you know.

[fortman]

n/m, wrapping the label_replace around the irate seems to work. Might still be useful, but I worked around it.

[wentwog]

I have such use-case: I have metrics like "command1_countRejected", "command2_countRejected", ...
For each command type, I want to see the maximum number of that command executions rejected within the last 30m, and draw it all as a rolling graph. I.e. use a query like
max( max_over_time({__name__=~".*countRejected"}[30m]) ) by (__name__)
to get
But since __name__ is lost at {__name__=~".*countRejected"}[30m] step, I can't get stats per each command, only cumulatively.

[deeTEEcee]

working with rabbitmq + kubernetes and a specific queue, the pod names are switching. I'd like to use avg_over_time with a certain time period but that leads to a result dependent on how many times the pod switched, so id like to ignore that pod name somehow.