Hi @Clara,
You're absolutely right, with the rarefying technique you are tossing away the majority of your hard-earned reads, in fact that is one of the major cited points against rarefying. However, as discussed in the papers that was linked in the other thread, the choice of rarefying or normalization seems to depend on the data itself. When I suggested that 6,000 is sufficient that is based mainly on my own experience and what I have seen others use/report. The paper on the other thread compared rarefying 3,000 reads/sample to other normalization techniques and found it to be comparable in most cases, so I would say 6,000 reads for human data is likely sufficient still. This is very much so an active topic of research and there aren't yet any standard rules and techniques to deal with it across all cases. That being said, some available techniques such as ANCOM and gneiss, use compositional data instead of raw counts and so circumvent -for the most part- the issue of uneven sampling depth. You would still remove low counted samples in those analyses though. For other analyses such as alpha & beta diversity the reads still need to either be normalized or rarefied. Currently, only the latter option is available in qiime2, but I imagine normalization techniques will arrive soon enough here as well. Hope that helps a bit!