This is a Windows-specific problem. After hard crash (actually I intentionally pulled the power plug on the test machine), MongoDB service got stuck in the "Starting" phase infinitely trying to start complaining about "old lock" (yes, the journalling is also turned off). I suggest there should be either a limit for autorestarts (I ended up with several gigabytes worth of logfiles filled with "start, oh-there's-an-old-lock-file, stop") on such failures or no autorestarts at all, at least for problems that aren't going to solve themselves. The existing SysV init.d script behavior looks fine to me, autorestarting on failing to obtain a lock or bind to an interface doesn't look to improve robustness under Windows, especially getting out of the "Starting" state ain't that easy.