You nailed the benchmarking gap. I've seen teams burn weeks trying to prove "improved performance" during the correlation period, only to realize their proxy overhead wiped out any gains from the new vendor's faster query engine. The goal has to be functional equivalence, not performance, or you're setting yourself up for disappointment.
Totally agree on the sunk cost trap for data transfer. If it's not a line item, it gets buried in "engineering innovation time" and the ROI model quietly dies. You need to amortize those developer hours against the vendor savings, and suddenly the three-year payback period stretches to five.
Keep automating!
The proxy approach can definitely simplify schema headaches, but it adds a hidden operational layer. We learned this the hard way when our proxy's memory usage spiked during a traffic surge, which we only caught because we'd instrumented the proxy itself with the new observability stack. Are you monitoring the proxy's own health and performance metrics? If not, that black box could fail silently during your critical correlation period.
Also, on the S3 budget, was the $1.2k/month target based on your old vendor's data granularity? Open-source stacks can be more verbose. We had to implement aggressive log filtering and metric downsampling policies after the fact to stay within budget, which later complicated some historical trend analysis.