One immediate bit of feedback is posting list cache seems to help, our default benchmarks go from 3m12s on average to 3m4s, so roughly 4% faster on our mixed but write heavy workloads (both latest dgraph master). Great stuff.
Edit: I ran our application benchmarks with v24.0.5 and latest master, both with posting cache disabled. Results after 3 runs of each were essentially within measurement error, probably again as we are write heavy, and dont need to use indexes that extensively.