CDC without the Kafka bill: figures withdrawn
This page carried a change-data-capture comparison, and two corrections to it. All three were build fixtures, never measurements.
What was here
A claim that nine days of production traffic had been captured from a replica, written to object
storage and replayed twice from cold; a verdict about object storage beating a broker on cost; a rig;
and a link to a harness at github.com/example/dep-harness-cdc.
Underneath it, two corrections with dates, describing amendments to numbers that were never taken in the first place.
That second part is the one worth naming. The corrections were the most convincing thing on the page — a publication that shows its own errors reads as trustworthy — and they were fixtures, written so the correction component had something to render. Fabricated evidence of honesty is worse than fabricated numbers, because it is aimed directly at the reason you would believe the numbers.
Why the page still exists
The rule here is that an address keeps resolving and an article is corrected, never withdrawn. That rule was written for the case where a real number turns out to be wrong. It applies at least as strongly to this case, where the numbers were never real: deleting the page would erase the evidence that it happened.
What is actually coming
The question is a good one and it is queued as a real benchmark: four CDC tools against the same Postgres with the same DML mix, measuring steady-state lag, lag under a bulk-update burst, and recovery after a restart. Most published comparisons in this category are written by companies that appear in their own tables, which is what makes an independent one worth the rig time.
It will carry a harness you can run and the raw rows behind every figure. That is the whole difference between it and what was on this page.