CDC's current algorithm to create a span-based checkpoint (which is reconstituted as a starting point for fine-grain checkpointing during replanning) works well if there is a single lagging span, but is naiive in situations where there is a spread between spans' latest resolved timestamp or multiple lagging ranges. We could significantly reduce duplicate emissions and improve changefeed performance by even marginally improving on the fine-grain checkpointing algorithm.
CDC's current algorithm to create a span-based checkpoint (which is reconstituted as a starting point for fine-grain checkpointing during replanning) works well if there is a single lagging span, but is naiive in situations where there is a spread between spans' latest resolved timestamp or multiple lagging ranges. We could significantly reduce duplicate emissions and improve changefeed performance by even marginally improving on the fine-grain checkpointing algorithm.
Jira issue: CRDB-41651