We experienced some data loss last month due to invalid serialization of log into the graph database.
That's why dogfooding at the core is our best course of action. We cannot publish our own success if our products fail XD.
We fixed it though and a ton of other bugs using careful auditing including reviews from LLMs. Many tools are involved like ASAn and top.
We debug in production and try to fix problems as they come up and by order of complexity. Sometimes earlier simpler problems resolution gives a way for simplifying the next problems that come up.