Notice that each append response, causes an update on materialized offsets, calling WiredTigerKVEngine::setLastMaterializedLsn, which then proceeds to perform a __conn_reconfigure. This accounts for roughly 2% of CPU. Why is this? (This is on the callback thread, so it slows everything down).
Note also that a read causes a flush(), so if the callback thread is delayed, previous writes might be sitting in the buffer?

See deep dive notes.