Test::Builder1.5 is slow. How slow? About 3x slower than 0.98. Enough to
significantly impact test performance. For example, Spiffy takes 1 second
with 0.98 vs 3.3s with 1.5.
I deliberately didn't do any profiling or optimization while the design was
coming together. This avoided spending time optimizing something that might
just be thrown out, and it focused on getting things working. And while my
policy has always been that performance is not a priority, a 3x drop is
unacceptable.
use Test::More;
pass() for 1..10000;
done_testing;
0.98
real 0m0.747s
user 0m0.726s
sys 0m0.019s
1.5
real 0m2.301s
user 0m2.271s
sys 0m0.027s
One thing which I know for sure is slowing things down is its use of a pure
Perl Mouse::Tiny. Test::Builder1.5 has a copy of Mouse::Tiny bundled as
TB2::Mouse. This both avoids a dependency on Mouse and protects Test::Builder
against Mouse bugs.
Test::Builder 0.98's performance benefits from being a monolithic object which
can quickly grab at it's own guts. Test::Builder1.5 is well factored out into
objects with attribute accessors. Pure Perl accessors can't compete with hash
lookups.
But Mouse boosted with XS can.
1.5-gonzales
real 0m1.702s
user 0m1.675s
sys 0m0.025s
Just switching over to my installed Mouse with XS shaved off a big chunk of
the runtime. Enough that I'm confident profiling can knock off the rest,
there's a lot of easy fat in there. I've already lopped it down to 1.368s
with two small changes.
The idea would be to continue to ship a copy of Mouse::Tiny as TB2::Mouse,
that takes care of the dependency loop, but to only use it if a good enough
version of Mouse is not already installed.
To mitigate risk of a new version of Mouse breaking Test::Builder, I'd ask the
Mouse folks to test their releases against Test::Builder1.5. I think they're
already doing that.
Thoughts?
--
60. "The Giant Space Ants" are not at the top of my chain of command.
-- The 213 Things Skippy Is No Longer Allowed To Do In The U.S. Army
http://skippyslist.com/list/
