Varnish: How to cache for failure, always ask backend if healthy?

I want to configure varnish (still on version 3) in the following way:

If the backend is available, every request should go to the backend, so no cached results should be used. If the backend is unavailable (so not healthy), I want to use the cached results for the next day. If after one day the backend is still unhealthy, an error should be shown.

I tried the following configuration using the grace mechanism, but this doesn't work: Although all requests are going to the backend (so no cached answers if the backend is available), in the case of a failure, I instantly get an Varnish Error (Unavailable) instead of Varnish using the cached data. What am I doing wrong?

sub vcl_recv {
  if (req.backend.healthy) {
    # No caching if backend is healthy
    set req.grace = 1s;
  } else {
    # Otherwise accept requests up to 1 day old
    set req.grace = 1d;
  }
}

sub vcl_fetch {
  # Cached requests are not valid at all
  set beresp.ttl = 1s;

  # Keep requests for ttl + 1 d
  set beresp.grace = 1d;
}