When we crash in the logging subsystem, we don't send crash logs back through that same subsystem. Instead, we write them directly to stderr (see logNoRecursion and printStackTraceNoRecursion in signal_handlers_synchronous.cpp).

When we start the server with --fork, we connect stdin/stderr/stdout to /dev/null (see forkServer() in initialize_server_global_state.cpp). So, these failures go nowere. We might say "we get a coredump, isn't that good enough" but we don't necessarily get a coredump depending on the system's configuration. For example, atlas might configure a limit of 0B for coredumps, so in fact we get nothing but a dead server.

Also, third-party libraries like OpenSSL might write errors to stderr, and we lose those too.