WarpStream's Kafka producer proxy holds each response to WarpStream's own write latency, so a mistuned producer fails during the proxy phase instead of after the cutover

Sources: WarpStream blog post, HN discussion
Summary: During migration the proxy delays each produce response to match WarpStream's own write latency, so a producer with timeouts too tight for the destination fails while the source cluster is still authoritative. The cutover gate uses time lag rather than offset lag, because time lag translates directly into how long producers sit retrying with writes held in client buffers. A topic that does not drain inside the timeout rolls back to the proxy state on its own, and the post is written by the vendor selling the destination cluster.
Why it matters: Gating the cutover on time lag rather than offset lag measures the quantity that decides whether producers stall, which is how long the backlog takes to drain rather than how many records it holds.