It seems that aggregate can occasionally cause a timeout with a tcp error like
unable to decode message length: read tcp 10.1.129.128:65170->3.216.165.122:27017: read: operation timed out
even when no options related to timeouts have been set.

This operation should not timeout, as the default socket options are supposed to block indefinitely. Some timeout, either default or manually-set, is causing this failure.