Yep so it definitely seems like either the memory or the 0.5 vcpu are a bottleneck -- g1-small isn't listed on the google cloud marketplace set up (and for this reason), but I thought I'd try and cut some costs and manually change to g1-small.
I'll stick with the n1-standard 3.75 machine type.
Still curious about why there is a 3x difference in discourse vs actual memory requirement, but not a big deal