Hi all! I came across FLINK-38349 while browsing JIRA for something to contribute to. Since it was marked as a blocker, I wasn’t sure how urgent it actually was, but I went ahead and reproduced the original issue and put together a fix. I've left a good bit of notes in both the ticket and the proposed PR. The patch itself resolves the original floating-point precision bug (which was resulting in zero-division exceptions) and includes two small gists demonstrating correctness (with the pre-bug behavior) along with another documenting performance numbers with the new approach (~4× improvement). Just sharing here for visibility in case the "blocker" tag reflects a real urgency.