Thanks a lot @LarsStegman for your investigation!!!

How about moving the copy into the running_output model and pass in a flag if a copy is required like

	for metric := range unit.src {
		for i, output := range unit.outputs {
			output.AddMetric(metric, i < len(a.Config.Outputs) - 1)
		}
	}

and in the model do

func (r *RunningOutput) AddMetric(m telegraf.Metric, requireCopy bool) {
	metric := m

	ok, err := r.Config.Filter.Select(metric)
	if err != nil {
		r.log.Errorf("filtering failed: %v", err)
	} else if !ok {
		r.metricFiltered(metric)
		return
	}

	if requireCopy {
		metric = m.Copy()
	}

	r.Config.Filter.Modify(metric)
	if len(metric.FieldList()) == 0 {
		r.metricFiltered(metric)
		return
	}
        ...
}

This way we do not need to expose the interna of the output model into the agent.

Yeah, that's also a good solution for me. I will make that change!

Download PR build artifacts for linux_amd64.tar.gz, darwin_arm64.tar.gz, and windows_amd64.zip.
Downloads for additional architectures and packages are available below.

👍 This pull request doesn't change the Telegraf binary size

@srebhan I am not sure why the memory leak test is failing. I should be making fewer allocations, not more. Do you have any idea?