I publish occasional notes on quantitative finance, analytics platforms, and engineering trade-offs.
Why shared_ptr is the wrong layer for objects with persistent identity and cross-language lifetime, and how Quantara’s handle registry - generational indices, lock-free refcounting packed into one atomic word, four typed wrappers - solves the problem properly.
This is a follow-up to Why Large Quantitative Analytics Platforms Rarely Fail All at Once. That post argued platforms slowly lose coherence; this one is about one structural choice that helps keep it. Thesis: When dependencies are knowable (as they often are in pricing and risk), explicit orchestration plus explicit caching scales better than runtime discovery plus database-as-memoisation (i.e., when the persistence layer quietly becomes the cache and dependency catalogue). This is not an argument against graphs. It is an argument about where you pay complexity. ...
Why do large quantitative analytics platforms fail? The unsatisfying but honest answer is: it depends. There is rarely a single root cause. Platforms almost never collapse because of one flawed model, one architectural mistake, or a single ill-judged rewrite. Most platforms do not break suddenly. They slowly lose coherence. They continue to run, produce numbers, and deliver outputs, yet over time the relationship between assumptions, inputs, and results becomes increasingly difficult to explain. Understanding this distinction matters. There is a difference between a system that stops working and a system that no longer makes sense. ...