Prometheus 在不使用组语句的情况下在加入请求中添加额外的 label？

Question

I am working with Prometheus. I have the following metrics:

Metric1

{"metric":{"namespace":"ns1", "instance_id":"someid", "label2":"somevalue", "label4":somethingelse"},"values":[[1675145655,"1"]]}

Metric2

{"metric":{"namespace":"ns1", "instance_id":"someid", "label3":"someothervalue"},"values":[[1675145655,"1"]]}

So both metrics have namespace and instance_id label, and then they have additional labels (label2, 3, 4).

I am doing a join request on them with the following request:

(metric1{namespace="ns1"} == 1) * on (namespace, instance_id)  (metric2{namespace="ns1", label3="something"} == 1)

Now, this is 1:1 join, for a specific time, I should never have two metrics with the same instance_id, so no need to perform many_to_1 or many_to_many here.

However, I wish to include in the result label present in metric1, but not metric2. If I do the following:

(metric1{namespace="ns1"} == 1) * on (namespace, instance_id) group_left() (metric2{namespace="ns1", label3="something"} == 1)

It works well, my result will contain the extra label from metric1:

{"metric":{"label2":"somevalue", "label4":"somethingelse"},"values":[[1675145655,"1"]]} 

What I am trying to confirm is:

• From my understanding, group_left/right command is made for many:many. I guess that since 1:1 is included in many:many, group_left/right also works for 1:1. But, can I add extra label in a simple 1:1 request? Or do I have to use the many:many join if I want to do so?

-On top of that, when I read documentation , it states that

Label lists can be provided to the group modifier which contain labels from the "one"-side to be included in the result metrics.

But here this the reverse I have, the label I get in my results are the one I did not put in the group_left. So for example if I do this request instead:

(metric1{namespace="ns1"} == 1) * on (namespace, instance_id) group_left(label2) (metric2{namespace="ns1", label3="something"} == 1)

Then the result give me metrics like this:

{"metric":{"label4":"somethingelse"},"values":[[1675145655,"1"]]} 

I have all extra label BUT the one I put my group_left statement. Is that expected? Am I mis reading the documentation here?

Thank you!

Answer 1

It is OK to use group_left() for 1:1 query if you need retaining all the labels from time series on the left side (or from the right side if group_right() is used). This doesn't introduce performance penalty.

If you put a list of labels inside group_left() , then these labels will be copied from time series on the right side additionally to all the labels from time series on the left side of the specified binary operator .

See more details in the official docs .

If you need retaining only some labels from the left side, then just drop unneeded labels after the calculations of q1 * on (...) group_left() q2 . Prometheus doesn't provide functions, which can drop unneeded labels from time series (or leave only the needed labels), but this functionality can be emulated with sum(...) without (...) or sum(...) by (...) in most cases. For example, the following query leaves only (namespace, instance_id, label2) labels from time series matching metric1{namespace="ns1"} in the result:

sum(
  (metric1{namespace="ns1"} == 1)
    * on (namespace, instance_id) group_left()
  (metric2{namespace="ns1", label3="something"} == 1)
) by (namespace, instance_id, label2)

Note that the sum() may return unexpected results if multiple time series have the same set of output labels specified inside by (...) .

PS Take a look also at label_del and label_keep functions provided by VictoriaMetrics - the Prometheus-like monitoring solution I work on. These functions are easier to use instead of sum(...) by (...) or sum(...) without (...) trick when some labels need to be dropped / left in the result.