Timescale engineer here. I'm betting we'll see a nice win; we tend to see write-mostly workloads the UNDO shouldn't be too expensive, and the smaller tuple sizes should be nice. We've built Timescale to be compatible with custom storage engines, so it should work as a drop-in, though of course until we've tested it we won't be sure.