A mistake not specifically related to PromQL, but easy to make due to the flexibility of the language, is to apply transformations to aggregate values, thus losing their statistical significance. As an example, you might be tempted to average pre-computed quantiles in summaries from a group of instances to get a feel for the cluster, but they can't be further manipulated from a statistical standpoint – the result from that operation would not resemble the corresponding quantiles of the cluster as a whole. This, however, can be done with histograms, since buckets from each instance can be summed cluster-wide before calculating the approximate quantile. Another common example of this is averaging averages.