Skip to content- Sources: primary, discussion
- Summary: The post describes replacing a linear subscriber filter in a critical message bus with an index keyed per topic partition, which is the part that shipped to production. The results section calls the 30 percent CPU reduction a preliminary staging measurement seen in some real-world scenarios, not a production result. A separate workstream on adaptive subtree-store splitting, motivated by an initial recovery that normally takes under 2.5 seconds taking over 13 minutes, is deployed only to staging.
- Why it matters: The linear filter was cache friendly and fast until client count grew, which is the failure mode where an algorithmically poor choice holds up until concurrency changes.