I merged this and announced the change on the development mailing list.

I handled your comments.

I left a few more comments.

Add TODO in summary field.

Thanks! I'm done handling your comments.

I left yet more comments, but they are only local, so we're getting there. :-) I
didn't check the tests, and I didn't think too hard about whether everything
will always work as intended. I suppose if it misbehaves in some way, we will
find out from the tests or from things going wrong in experiments soon enough. :-)

But with this sort of change, I would recommend writing an email to ai@... or
downward-dev@... after it has been merged explaining what was changed and why
and asking people to watch out for any odd behaviour over the next few weeks.

I handled your comments and think this is ready for another review.

Left some more comments.

Thanks! I took care of your comments.

I left some comments on bitbucket.

Taking our off-tracker discussion into account, my understanding is that we want to let the 
driver ignore external limits. The following pull request should do just that and I adjusted the 
issue title accordingly:

https://bitbucket.org/jendrikseipp/downward/pull-requests/107/

If the internal time limit surpasses the external hard limit, it is a failure and the driver 
aborts. 

Working on this I uncovered a subtle bug: the driver captures all stderr output when calling the 
translator, including the preexec function that is used to set the resource limits. If that 
function encounters an error (like trying to raise a hard resource limit), we print the error to 
stderr and exit the driver immediately. Since the stderr output is captured, however, it is 
never shown to the user.

The pull request addresses this in the last commit 
(https://bitbucket.org/jendrikseipp/downward/commits/426a105737dcd56ac8243a2eba06c33d7eb6ba5d). 
If we encounter an error in a child process, we use os._exit() to exit only the child process, 
not the whole interpreter. This requires passing around is_subprocess Booleansin quite a few 
places. If you have a more elegant solution, please let me know.

I think we introduced a regression in issue775:

$ ulimit -Sv 102400
$ ./fast-downward.py --translate ../benchmarks/gripper/prob01.pddl

results in 

[...]
INFO     translator time limit: None
INFO     translator memory limit: None
[...]

The change was made in the get_memory_limit() and get_time_limit() functions in 
driver/limits.py in this changeset: http://hg.fast-downward.org/rev/33275c9cc1cf

I think the fix for this issue is to only let these two functions return None 
only if all three limits (component, overall and external) are None.