I ran the test for a bit longer extending it 57k visits to see if it timing stays linear with visits. Here is a bunch of plots from this extended test.
The total per-visit time continues to grow afrer 30k visit but it is not very linear anymore (or it is linear but with different slope):
Select time still scales linearly, major contribution comes from selecting DIASources:
but insert time stays more or less constant after 30k visits, largest contribution to insert time comes from DIAObject:
That time stays about constant after 30k, this can be explained by the number of inserted DIAObject as it should stabilize after 30k as we have a cutoff of 30 days for forced photometry of "noise" objects (at least this is what is implemented in my prototype).
DIASource select is clearly still a major concern as it will continue to grow until we reach 12 months of history cutoff. Significant insert time is likely a limitation of cloud storage.