All my computers thank you greatly (especially my poor EeePC). I haven't tested this yet, though, but this has been a thorn in my side for a long time. Tests take a long time for me to run because my computer starts thrashing the swap-space, and then cleaning out the virtual memory at the end of the tests takes a long time.

Thanks!

any objections to merging this?

apologies, i pushed the comment and close button instead of simply comment

Paul, I think this has gotten too complicated, so I don't recommend merging this changeset. Instead, I recommend that your close-fig branch, or some variation of it, be applied to v1.0.x and merged into master ASAP. Failing to close the figures in the tests is a real bug, and should be fixed in the maintenance branch. There is another major bug in testing in master that I would like to fix, but to avoid more complicated merges, I would like to see the close patch applied first.
Resetting the subplot parameters is a separate issue, and should not get scrambled up with the basic bug fix.
Regarding that, I agree with Jouni that an rcParam is overkill; the rc system should be used sparingly. I get the impression that all we need may be a figure method that resets subplot params.
The value of reusing a figure in testing sounds marginal at best; but if it is to be done, it should be a separate changeset in master.

More miscellaneous comments:
It is not at all clear to me that a user would expect or want clf to reset the subplot params; and if a user did, it would not be obvious that that is what a kwarg called "scrub" is for. My guess is that in most cases where a user really wants to reuse a figure, the user will not want to have the subplot params reset. Better to leave clf() simple, let it just continue to delete artists. If one really wants a clean slate, one can start with a new figure--and even then one has to be careful because there are other ways that global state can be altered. Not just via the rc system, but via units support, and probably other ways I haven't thought of. (The units support is currently fouling up a subsequent test; this is what I want to get fixed as soon as the close/clf issue is resolved.)
The reason close() is slow is that it is calling gc.collect() to ensure cyclic references are cleaned up immediately. Maybe this is overkill, but I suspect it is better left in place than removed or made optional. This might be worth some investigation.
We can probably save 30-40 seconds on the tests merely by modifying the hexbin test; again, I want to do that after close/clf is resolved.
I suspect it would also be possible to speed up the tests a bit by modifying the image_comparison decorator so that it takes a function returning a figure, and then uses the savefig method on that figure to generate the various file types. That would cut the number of figure generations down by a factor of 3. That decorator is horrendously complicated, though.

After thinking on this for a while, I think I agree. I think fixing the
memory consumption in the tests is important, but the public API should be
driven by the needs of general usage. In that latter context, the "scrub"
kwarg doesn't seem terribly useful.

But let's get all those figures closing in the tests come hell or high water
-- the difference is staggering!

On Sun, Apr 24, 2011 at 7:10 PM, efiring <
reply@reply.github.com>wrote:

I suspect it would also be possible to speed up the tests a bit by
modifying the image_comparison decorator so that it takes a function
returning a figure, and then uses the savefig method on that figure to
generate the various file types. That would cut the number of figure
generations down by a factor of 3. That decorator is horrendously
complicated, though.

That's a good idea.

I have some time today -- I'm going to look into cleaning up that decorator and see if we can save the figure regeneration time -- it's also a good place to automatically close figures when done which probably addresses the problem this pull request addresses.

ok, i'm closing this one in favor of #96