1. 4

  2. 2

    The low GC grace period would concern me a bit until they upgrade Cassandra. If there is any kind of operational issue that prevents repairs from finishing for 2 days then they could have some resurfacing data. But deleting lots of data is a bad experience in Cassandra.

    1. 3

      It becomes a cost/benefit thing. How bad is it for Discord if messages come back? Not terrible. It’s not like these are financial transactions, which we emphatically don’t want replayed. Also, it’s relatively easy to say, “if the repairs don’t complete within our GC grace period, we’ll wipe the node and rejoin it as a fresh instance.” The cost there is a pile of row migration, which isn’t great, but also is not terrible.