The devil will absolutely be in the details, but the speaker/author identifies a real pain point that bolting on a replication-log tailing system to provide realtime capabilities results in numerous write bottlenecks unless the system is architected from scratch to avoid them.
Meteor has the op-log-driver which calculates query-result for changes instead of running the wohle query again. Same comes with rxdb where it's called query-change-detection.
That's true, but monitoring the stream of change events in a single machine (be it server or client) makes this machine your bottleneck for overall system throughput.
And the paper for their more scalable alternative: https://www.cl.cam.ac.uk/~mks40/pubs/vldb_2017.pdf
The devil will absolutely be in the details, but the speaker/author identifies a real pain point that bolting on a replication-log tailing system to provide realtime capabilities results in numerous write bottlenecks unless the system is architected from scratch to avoid them.