Your PR seems to leak memory. I have a working patch for the related issues which does not allocate dynamic memory. I however need to fix the tests.

IMO not freeing heap-allocated memory is fine as long as it can be accessible via a global pointer. Tools like valgrind should categorize this as "still reachable" not a leak. I agree it'd be nicer we can get rid of it though. I'm looking forward to your patch, thanks!

$ cat still_reachable.cc
__thread int* global;
int main() {
global = new int();
}
$ g++ still_reachable.cc
$ valgrind --leak-check=full ./a.out
...
==137829== HEAP SUMMARY:
==137829== in use at exit: 4 bytes in 1 blocks
==137829== total heap usage: 1 allocs, 0 frees, 4 bytes allocated
==137829==
==137829== LEAK SUMMARY:
==137829== definitely lost: 0 bytes in 0 blocks
==137829== indirectly lost: 0 bytes in 0 blocks
==137829== possibly lost: 0 bytes in 0 blocks
==137829== still reachable: 4 bytes in 1 blocks
==137829== suppressed: 0 bytes in 0 blocks

What if your glog client application create and destroy 1000 threads (because it is supposed to work continuously for 2 years and is creating 3 threads per day for various tasks) ? At some point of time your pull request will bring a trouble(sooner or later) and not freeing heap allocated memory in C++ is real issue for every application.

Added #226.