When a service fails to start, a retry is done after 100ms for a total of 5 tries, after which it's not tried automatically again. Does this make sense?

If a service crashes (for the first time) it might not make sense to wait 100ms.
If a start attempt doesn't succeed, it might make sense to wait more than 100ms. Perhaps some exponential backoff should be used up to a minute or so, instead of not trying at all after about 1s.

215-17+deb8u7 @ Debian 8