Hello Ivan,
Is there any chance that this is a database I released (unite-train)?
If so, I trained it to include all sh_ suffixes for species hypotheses (SH).
Having all level 7 labels be unique increases memory usage a lot!
In my new version, I remove the species hypotheses suffix and it needs less RAM to run. If you are using my old database, try my new database:
This particular Qiime2 plugin is could be improved, see Utilize one database instance for multiple classification jobs · Issue #190 · qiime2/q2-feature-classifier · GitHub
If you have the python skill, take a look!