Comparative benchmark modes--apples to apples (by params) and everything vs. one trial This will help guide optimization work
Figure out where the huge spikes in operation time come from (tiered resizing?)
We need a more thorough set of tests for the messaging system to saturate deletes, to ensure we don't have more lurking bugs
Implementations of backwards scan, pred, and succ
Write the WAL, back with in-mem and redis, then add to benchmarks. This is necessary for disk-backed work
Choose splits and sizes based on serialized results--big perf gain
benchmark dataset like (map + (repeatedly rand) (iterate #(+ 0.01 %) 0.0)) that's a sliding random window, for some random moving write heavy region with lots of cold nodes
Add async writes
Make a snapshot life-extender
Add support for datascript