Interned strings are immortal, despite what the documentation says

[encukou]

Bug report

Bug description:

The sys.intern documentation explicitly says:

Interned strings are not immortal; you must keep a reference to the return value of intern() around to benefit from it.

However, they were made immortal in #19474 (Implement Immortal Objects -- implementation of PEP 683), without a documentation update. The 3.12 What's New also doesn't mention the change. [edit: it's now clear that the PR author (@eduardo-elizondo) intended it but the reviewer (@markshannon) did not]
PEP-683 itself only mentions the change as a possible optimization.

Meanwhile, pathlib interns every path segment it processes, presumably depending on the documented behaviour. In CPython's test suite, that causes names of temporary directories to leak (stealthily, since interned strings are excluded from the total refcount).
[edit: Worse, type_setattro, PyObject_SetAttr or PyDict_SetItemString immortalize keys, so strings used for key/attribute names this way are not reclaimed until interpreter shutdown.]
On the other hand, free-threading (PEP-703) seems to rely [edit: relies] on these being immortal.

What's the correct behaviour? [edit: specifically, in 3.12 and non-free-threading 3.13, should this change be reverted or documented?]

@eduardo-elizondo @ericsnowcurrently

Linked PRs

• gh-113993: Allow interned strings to be mortal, and fix related issues #120520
• [3.13] gh-113993: Allow interned strings to be mortal, and fix related issues (GH-120520) #120945
• gh-113993: For string interning, do not rely on (or assert) _Py_IsImmortal #121358
• gh-113993: Don't immortalize in PyUnicode_InternInPlace; keep immortalizing in other API #121364
• [3.13] gh-113993: For string interning, do not rely on (or assert) _Py_IsImmortal (GH-121358) #121851
• [3.13] gh-113993: Don't immortalize in PyUnicode_InternInPlace; keep immortalizing in other API (GH-121364) #121854
• [3.12] gh-113993: Make interned strings mortal (GH-120520, GH-121364, GH-121903, GH-122303) #123065
• [3.12] gh-113993: For string interning, do not rely on (or assert) _Py_IsImmortal (GH-121358) (GH-121851) #124938
• gh-113993: InternalDocs: Add String Interning to README #127250

Related fixes:

• gh-121863: Immortalize names in code objects to avoid crash #121903
• gh-122291: Intern latin-1 one-byte strings at startup #122303
• gh-122420: Fix accounting for immortal interned strings in refleak.py #122421

[nascheme]

I would suggest that using "immortal" here could be confusing. The flagged checked by _Py_IsImmortal() is low-level implementation detail and should not be something that the library documentation discusses. Strings that you call intern() on may or may not have this flag set. Different Python runtimes might implement intern() some other way. The behavior that matters is that the string gets stored in an interned strings table and that future calls to intern() with the same string content will return an object with the same id(), even if the user program doesn't keep a reference to the string.

[encukou]

The behavior that matters is that the string gets stored in an interned strings table and that future calls to intern() with the same string content will return an object with the same id(), even if the user program doesn't keep a reference to the string.

No. That is not the behaviour on Python 3.11.
I get:

$ python3.11
Python 3.11.7 (main, Dec 18 2023, 00:00:00) [GCC 13.2.1 20231205 (Red Hat 13.2.1-6)] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import sys
>>> id1 = id(sys.intern('some string'))
>>> other = sys.intern('other string')
>>> id2 = id(sys.intern('some string'))
>>> id1 == id2
False

IMO, the guarantee is that if two strings are interned, they can be compared by identity.
CPython uses this for a fast path in string comparison; so some users call intern on strings that are likely to be compared to one another -- even ones that srent expected to live until interpreter shutdown. We can say that's optmizing for a specific implementation, but it's common -- even pathlib and the compiler (e.g. eval) do it.

The new behaviour can leak strings that are not likely to be used again.

[erlend-aasland]

$ cat test.py
import sys
id1 = id(sys.intern('some string'))
other = sys.intern('other string')
id2 = id(sys.intern('some string'))
assert id1 == id2
$ python3.8 ./test.py
$ python3.9 ./test.py
$ python3.10 ./test.py
$ python3.11 ./test.py
$ python3.12 ./test.py
$ python3.13 ./test.py

No assertions triggered.

[encukou]

The the new string is allowed to reuse the old one's memory; I guess this is build-specific. What platform are you on?
Instead of other = sys.intern('other string'), try putting in gc.collect and something more intensive, like import asyncio?

[erlend-aasland]

I'm on macOS using the official build. Strange thing: I can reproduce the behaviour you're seeing in the REPL.

[erlend-aasland]

Instead of other = sys.intern('other string'), try putting in gc.collect and something more intensive, like import asyncio?

Nope, cannot reproduce using $ python3.11 ./test.py as above.

[encukou]

Ah! Right, literal constants are part the code object, which does normally stay around. You'll need something that's not folded to a constant, like:

$ cat test.py 
import sys
def intern_and_report(string):
    string = sys.intern(string)
    the_id = id(string)
    print(f'{string!r} at {id(string)}: refcount {sys.getrefcount(string):x}')
    return the_id

id1 = intern_and_report('a' + __name__)
str_b = sys.intern('b' + __name__)
id2 = intern_and_report('a' + __name__)
assert id1 == id2

$ python3.10 ./test.py
'a__main__' at 139775546163056: refcount 3
'a__main__' at 139775546294640: refcount 3
Traceback (most recent call last):
  File "/tmp/rttair/./test.py", line 11, in <module>
    assert id1 == id2
AssertionError

$ python3.12 ./test.py
'a__main__' at 139964910181616: refcount ffffffff
'a__main__' at 139964910181616: refcount ffffffff

[erlend-aasland]

Yes; thanks!

[erlend-aasland]

See also #113601

[eduardo-elizondo]

@encukou just getting to this! You're right, as of now, all interned strings should be immortal, so the sys.intern docs should probably be updated to update that they are indeed immortal and the fact that we don't need to keep a reference to benefit from it anymore.

As for the leaks, as @erlend-aasland already called out, we have a PR to correctly clean them up in #113601!

[nascheme]

The behavior that matters is that the string gets stored in an interned strings table and that future calls to intern() with the same string content will return an object with the same id(), even if the user program doesn't keep a reference to the string.

No. That is not the behaviour on Python 3.11.

Sorry, I was not very clear. I was describing the behaviour we have with Python 3.12 and we had with older versions of Python (version 2.2 and before, it seems). It looks like this commit changed it: 45ec02a.

Since 3.12 changed this, definitely the documentation should be updated. Further, I think we should have had a discussion and explicitly decide if this is desirable behaviour, since it reverts the effect of 45ec02a. Maybe there was one and I missed it? Perhaps free-threading doesn't require that every interned string become immortal but only some of them (e.g. symbols used by code objects etc). That might preserve free-threaded performance (e.g. interned strings shared between threads) while avoid the "leak" like encountered by pathlib (e.g. interned strings only used for a short time and then discarded).

[encukou]

@eduardo-elizondo

as of now, all interned strings should be immortal

To be clear, I think this is bad. Some interned strings are temporary, and they now leak. They should be cleaned up sooner that at interpreter shutdown.

[jvs]

The documentation for sys.intern explains, "Interned strings are not immortal; you must keep a reference to the return value of intern() around to benefit from it."

Some open source libraries depend on this behavior, like msgpack-python. Changing the behavior of sys.intern will cause problems for some applications. I work on an application that is experiencing out-of-memory errors because of this issue.

One solution might be to add a separate function called sys.immortalize(string). Alternately, an optional argument to sys.intern could also do the trick. For example, sys.intern(s, immortalize=True).

Restoring the old, documented behavior of sys.intern would save people a lot of trouble. Making the new behavior opt-in (with a separate function or an optional argument) might be a way to have our cake and eat it too.

• https://docs.python.org/3/library/sys.html#sys.intern

[jvs]

Does this mean that PyDict_SetItemString will also immortalize its key argument? Will every key passed to PyDict_SetItemString live forever?

It seems like that could also be a source of leaks. Maybe the documentation for PyDict_SetItemString should be updated to indicate that key will be kept alive forever.

Alternately, maybe PyDict_SetItemString should not call PyUnicode_InternInPlace. I notice a comment on that line asks, XXX Should we really?. Seems like a good question.

• https://github.com/python/cpython/blob/main/Objects/dictobject.c#L4926
• https://docs.python.org/3/c-api/dict.html#c.PyDict_SetItemString

[markshannon]

I don't see a good reason why interning strings should make them immortal.
The two things, immortality and interning, are separate concepts.

Why were interned strings changed to be immortal in 3.12?

Since interned strings that are tied to code objects are effectively immortal, making those actually immortal seems fine if necessary for performance.

But strings interned with sys.intern() should not be made immortal implicitly, IMO.
@jvs suggestion of adding an immortalize argument to sys.intern() seems a good one.

[markshannon]

The immortal objects PR seems to allow interned, but not immortals strings:
https://github.com/python/cpython/pull/19474/files#diff-6be7f081fe6c5e9cbc89323a00399b291d1cda855bcb4c6eeaee0fac89c2f8ddR105
so it is a bit surprising that they aren't supported.

[markshannon]

It might also be worth making the distinction between "interpreter immortal" and "process immortal" strings, as the latter can be shared between multiple interpreters.

[colesbury]

The free-threaded build needs every interned string to be immortal, not just symbols, if you don't want to introduce bottlenecks that prevent scaling in multi-threaded programs.

The documentation can describe the possible behaviors without promising a specific implementation. Something like:

"Depending on the Python version and implementation, interned strings may or may not be immortal; you should keep a reference to the return value of intern() around to benefit from it."

[markshannon]

Why do all interned strings need to be immortal?
Strings where there is potentially high contention on the refcount need to be immortal, to allow scaling.
But that only implies that interned strings must be immortal if there is a strong correlation between contention and interning.
Is there such a strong correlation?

Are there not strings that are contended and are not interned, or does the free-threading build intern so many strings that all possibly contended strings are interned?

[colesbury]

Interning implies sharing and sharing between threads leads to reference count contention. To take the pathlib example, I'd much rather pathlib did not intern strings, but if it does intern the components, it's important that those strings are immortalized, because otherwise you are likely to have a bunch of strings shared between threads that otherwise would not be shared.

Are there not strings that are contended and are not interned...?

It's definitely possible, but from what I've seen it's not the most common source of reference count contention. If you don't intern a string, it's much less likely to be incidentally shared between threads.

[jvs]

This seems like a breaking change that is kind of flying under the radar. The documentation explains, "Interned strings are not immortal; you must keep a reference to the return value of intern() around to benefit from it."

It looks like it been that way for 20 years

I'd be worried that there are many applications that depend on the old behavior. If there's a way to maintain backwards compatibility, while also finding a way to avoid bottlenecks around shared strings, then that might be the best option.

• 45ec02a

[methane]

Can we make it mortal again on --enable-GIL build?

[encukou]

Right. It looks like free-threading needs this, and the ecosystem (and stdlib) will need to adapt, so the question is how soon.
Specifically, what should we do in 3.12 -- revert the change, or document it?

I guess ultimately the RM should decide that.

[encukou]

The 3.12 backport is at #123065.

@Yhg1s, re:

Yes, I agree this should be reverted in 3.12, and conditional on free threading in 3.13. That's my opinion as RM, but considering the discussions the DC had about the immortal objects PEP, I'm sure they all agree.

Unfortunately it's far from a straightforward revert.
Do you still want this in 3.12? Should I ask the SC?