Currently the code does an explicit GC both before and after loading the gazetteer so that it can report on memory consumption. It does this even when duplicating a PR in which no extra memory is allocated. We've seen each GC take as much as 5s, so when using this on cloud or GCP it can add quite an overhead: 40s for 4 threads is just GC. On cloud this is possibly what's causing the prometheus warnings as the full load time is high.

The suggestion is to add a new debug or profiling option to the PR, so that in normal usage we don't report the memory consumption and don't explicitly run the GC.