How are you handling virtual threads in production so far?

We moved a fairly I/O-heavy service from a big fixed thread pool to virtual threads and throughput looks better, but debugging under load has been a different experience. We hit a few cases where synchronized blocks around blocking calls caused pinning before we updated our JDK and some libraries. Thread dumps also look very different with thousands of threads. For those running them in prod, what surprised you and what would you do differently?