Hi @zietho .
Thanks for your response. I am still a bit confused about your response.
To be honest, I was not getting the relevance of your example to obtain the mean of means completely.
I tried different queries, where i selected some metric, like `builtin:service.response.time` and `builtin:service.requestCount.total` to get for every service the response time and request count.
Then, I compared my computation in excel with the result dynatrace comes up with when adding various aggregations. I also tried what @zietho suggested for "get[ting] the mean of means".
That is I started with a query with:
builtin:service.response.time:filter(
and(
in("dt.entity.service",entitySelector("type(service),serviceType(~"WEB_SERVICE~")")),
in("dt.entity.service",entitySelector("type(service),tag(~"[Environment]app.tenant:xx~")")),
in("dt.entity.service",entitySelector("type(service),tag(~"[Kubernetes]app:yy-services~")"))
))
and a 2nd one where I use `builtin:service.requestCount.total` instead which also uses the same filter condition.
The resulting table has 3 columns, one for the service name, one for the response time and one for the request count.
I exported the table the DataExplorer to CSV for importing and checking the aggregation results.
In the follow-up changes I made to the response time query, I have disabled/removed the query for the request count, i.e. I only used the query for the response time metric.
Interestingly, and kind of strange is the Dynatrace query language because depending on the chaining of the same `avg` operators, you get a different result and semantics.
For your information:
- If you append to the above 1st query for the response time `:splitBy():avg` you will get the mean weighted by the request count. (as pointed out in my original post)
- But, if you instead append to the 1st query `:avg:splitBy()` only you get some sort of sum over the metrics. In my example you just get a single scalar value that is the sum of the response time. Strange behavior in my opinion, since neither `avg` nor `splitBy` does imply the meaning of a sum. 🤔
- However, when I tried to append only `:avg:splitBy():avg` to the query (as @zietho suggested), I ended up with the (unweighted) mean of response times. That is the sum of response times across web services (i.e. result of point 2) divided by the number of web services. Again, there is room for improvement with regard to query language syntax. 🤔 But, it seems this does not yield the mean of means in the selected time period.
I got these results by trying to attach these 3 different combinations separately and comparing them to the calculations with the table of response times and number of requests that I imported into Excel.
Concluding, the use of `avg` does not always mean take the (weighted/unweighted) average. 😖
Unfortunately, this is not reflected in the documentation -- yet.
Why is this important? Depending on your services' request count, there may be HUGE differences when comparing the mean and the weighted mean, if the request count distribution across services is skewed.
It would be nice if these findings would be somehow reflected in the Dynatrace documentation.
@zietho What is your opinion on that issue? Thank you for pointing me to try `:avg:splitBy():avg`.
Best regards,
Bernhard
