Currently, new_interpreter() is a subset of Py_InitializeFromConfig(): the code was duplicated. I would prefer that both functions stay in sync and so that new_interpreter() reuses more Py_InitializeFromConfig() code.

New changeset ef5aa9af7c7e493402ac62009e4400aed7c3d54e by Victor Stinner in branch 'master':
bpo-38858: Reorganize pycore_init_types() (GH-17265)
https://github.com/python/cpython/commit/ef5aa9af7c7e493402ac62009e4400aed7c3d54e

New changeset 7eee5beaf87be898a679278c480e8dd0df76d351 by Victor Stinner in branch 'master':
bpo-38858: Factorize Py_EndInterpreter() code (GH-17273)
https://github.com/python/cpython/commit/7eee5beaf87be898a679278c480e8dd0df76d351

> New changeset ef5aa9af7c7e493402ac62009e4400aed7c3d54e by Victor Stinner in branch 'master':
> bpo-38858: Reorganize pycore_init_types() (GH-17265)

This change introduced a reference leak:

https://buildbot.python.org/all/#builders/80/builds/771

test_atexit leaked [792, 792, 792] references, sum=2376
test_capi leaked [528, 528, 528] references, sum=1584
test__xxsubinterpreters leaked [17954, 17952, 17954] references, sum=53860
test_threading leaked [792, 792, 792] references, sum=2376

It should be fixed by PR 17286.

New changeset e7e699e4df73420ddccaa0057cd07ebb3b590b9b by Victor Stinner in branch 'master':
bpo-38858: Fix reference leak in pycore_init_types() (GH-17286)
https://github.com/python/cpython/commit/e7e699e4df73420ddccaa0057cd07ebb3b590b9b

New changeset fff7bbfdb6b7c143b73b6b4b6b40e828c101110c by Victor Stinner in branch 'master':
bpo-38858: Add _Py_IsMainInterpreter(tstate) (GH-17293)
https://github.com/python/cpython/commit/fff7bbfdb6b7c143b73b6b4b6b40e828c101110c

New changeset b93f31fcd9f10b213c614d4944baf9ca2df2016c by Victor Stinner in branch 'master':
bpo-38858: Fix Py_Finalize() when called from a subinterpreter (GH-17297)
https://github.com/python/cpython/commit/b93f31fcd9f10b213c614d4944baf9ca2df2016c

New changeset 5dcc06f6e0d7b5d6589085692b86c63e35e2325e by Victor Stinner in branch 'master':
bpo-38858: Allocate small integers on the heap (GH-17301)
https://github.com/python/cpython/commit/5dcc06f6e0d7b5d6589085692b86c63e35e2325e

> bpo-38858: Allocate small integers on the heap (GH-17301)

I associated this change to this issue because I had troubles when I tried to call _PyLong_Fini() in subinterpreters. I'm now trying to have different small integers per interpreter.

New changeset 3d4833488a173c16446c3f94f58f05e2d13c5dee by Victor Stinner in branch 'master':
bpo-38858: Call _PyUnicode_Fini() in Py_EndInterpreter() (GH-17330)
https://github.com/python/cpython/commit/3d4833488a173c16446c3f94f58f05e2d13c5dee

I introduced a workaround to a reference leak in bpo-36854 because _PyImport_Cleanup() doesn't clear modules dictionary in some cases:
https://bugs.python.org/issue36854#msg357160

Maybe _PyImport_Cleanup() should be enhanced to ensure that the dictionary of all modules is cleared when the function exit.

New changeset e0c9ab8e26d1648b870b80c296b2490a5e9553e5 by Victor Stinner in branch 'master':
bpo-38858: Add init_set_builtins_open() subfunction (GH-17346)
https://github.com/python/cpython/commit/e0c9ab8e26d1648b870b80c296b2490a5e9553e5

New changeset b00513636c9891deba4cae50217e25e8faf6c6ac by Victor Stinner in branch 'master':
bpo-38858: Add init_interp_main() subfunction (GH-17347)
https://github.com/python/cpython/commit/b00513636c9891deba4cae50217e25e8faf6c6ac

New changeset 82c83bd907409c287a5bd0d0f4598f2c0538f34d by Victor Stinner in branch 'master':
bpo-38858: _PyImport_FixupExtensionObject() handles subinterpreters (GH-17350)
https://github.com/python/cpython/commit/82c83bd907409c287a5bd0d0f4598f2c0538f34d

New changeset 2582d46fbcf7bdf86b9cf4016850b8d155267ac6 by Victor Stinner in branch 'master':
bpo-38858: new_interpreter() reuses pycore_init_builtins() (GH-17351)
https://github.com/python/cpython/commit/2582d46fbcf7bdf86b9cf4016850b8d155267ac6

New changeset 2ec1a1b307cc893adae4662a32e1d2e94b6908e3 by Victor Stinner in branch 'master':
bpo-38858: new_interpreter() uses pycore_init_import_warnings() (GH-17353)
https://github.com/python/cpython/commit/2ec1a1b307cc893adae4662a32e1d2e94b6908e3

New changeset 81fe5bd3d78f9bb955f8255404d99df27a31c36a by Victor Stinner in branch 'master':
bpo-38858: new_interpreter() reuses _PySys_Create() (GH-17481)
https://github.com/python/cpython/commit/81fe5bd3d78f9bb955f8255404d99df27a31c36a

New changeset d863ade0c7fa4826e8b71aa467809c83a711f019 by Victor Stinner in branch 'master':
bpo-38858: Add pycore_interp_init() code to factorize code (GH-17483)
https://github.com/python/cpython/commit/d863ade0c7fa4826e8b71aa467809c83a711f019

Py_NewInterpreter() (new_interpreter() in practice) and Py_InitializeEx() now share almost all their code.

Py_EndInterpreter() shares almost all its code with Py_Finalizer(). It's not perfect, but it's way better than previously.

Py_NewInterpreter() now isolates more things from the main interpreter. For example, builtins and sys modules no longer copy the module dictionary of the main interpreter, but create their own dictionary from scratch.

See each commit for the details.

Py_Finalizer() could share more code with Py_EndInterpreter(), but each Py_Finalizer() change is really tricky and require to pay a lot attention. The Python finalization is really fragile. I started to take notes on this fragile code:
https://pythondev.readthedocs.io/finalization.html

I consider that the initial issue is fixed, so I close the issue.

New changeset 080ee5a88406fb68aaab741145cd5d2a7c5f2ad6 by Victor Stinner in branch 'master':
bpo-38858: Fix ref leak in pycore_interp_init() (GH-17512)
https://github.com/python/cpython/commit/080ee5a88406fb68aaab741145cd5d2a7c5f2ad6

Thanks for working on this.  It really does have far-reaching benefits, not just for the subinterpreter stuff I'm interested in. :)

New changeset 630c8df5cf126594f8c1c4579c1888ca80a29d59 by Victor Stinner in branch 'master':
bpo-38858: Small integer per interpreter (GH-17315)
https://github.com/python/cpython/commit/630c8df5cf126594f8c1c4579c1888ca80a29d59

FYI this change broke librepo which calls PyLong_FromLong() without holding the GIL. In Python 3.8, "it works". In Python 3.9, it does crash: get_small_int() gets a NULL tstate and then dereference a NULL pointer.

librepo bug:
https://bugzilla.redhat.com/show_bug.cgi?id=1788918

IMHO it's a bug in librepo: the GIL must be held to use Python C API.