What did we do against such an unpredictable system? We took it apart: merging services that were too coupled; defining clear boundaries between services; moving some services away from the event-store into a traditional database; making some communication channels synchronous; writing integration and end-to-end tests.