Found the answer through the Slack user group, courtesy of an Adam Binford.

I had set `delta.logRetentionDuration='24 HOURS'` but did not set `delta.deletedFileRetentionDuration`, and so the checkpoint file still had all the accumulated tombstones since the table existed.

Since I was running a compactor every 15 minutes, the table itself (and thus the checkpoint file) would not consist of too many files, however since all the tombstones of every microbatch streamed in still existed, it allowed the checkpoint file to balloon in size. Once setting it to a lower interval, my batch time decreased from 20+ seconds down to about 5.