There are a couple of documents and articles online that suggest turning off read repair for tables using the time window compaction strategy (TWCS), as it mixes data that should belong to old buckets into new ones, which leads to higher read amplification.
I wonder if some common node operations will cause similar issues. Specifically, adding a node, replacing a dead node, running nodetool repair, running nodetool decommission or running nodetool rebuild.
Also, has ScyllaDB implemented some optimizations regarding TWCS compared to Cassandra?
All node operations, as well as repair, will split data into buckets appropriate for the TWCS parameters, so neither streaming, nor repair should cause any mixing of data from different buckets. You can repair your TWCS table, and you can add/remove/rebuild nodes, it shouldn’t cause any mixing of old/new data.