Implement support for DLR get/set by filmor · Pull Request #2706 · pythonnet/pythonnet

filmor · 2026-04-18T13:50:42Z

Implements #72 and should thus fix #2696.

@spkl Could you try out this branch?

spkl · 2026-04-20T12:17:55Z

@filmor If I understand this correctly, one wouldn't even need the dynamic.py script anymore for DynamicObject access to work, correct? That seems great :)
To test this in our setup, can you point me to a wheel that I can install, or provide documentation how to build it?

filmor · 2026-04-20T13:27:34Z

@spkl Yes, the dynamic support would be integrated. To build a wheel, the simplest way is to get uv and the .NET SDK installed on your machine, check out the repo and run uv build. That builds the wheel in dist/.

spkl · 2026-04-20T13:29:52Z

OK so I did a little digging and found out that I can pip install . in the root directory of the pythonnet repository to install this version into my venv.
Edit: Is this wrong? Should I try again with uv?

Getting the value of dynamic properties seems to work, but I encountered a problem when setting dynamic properties. This is what I observed when running a script like this:

x = obj.MyProp
obj.MyProp = None
y = obj.MyProp

For line 1, TryGetMember is called for name "MyProp" and x will correctly be assigned the value returned from there.
For line 2, TrySetMember is NOT called.
For line 3, TryGetMember is NOT called, y will be assigned the value from line 2 (None), but that value is unknown to the .NET implementation.

filmor · 2026-05-06T08:44:14Z

@spkl Could you test again whether this works for you?

spkl · 2026-05-06T19:52:25Z

Will do ASAP and let you know

spkl · 2026-05-07T08:46:52Z

OK @filmor, this is pretty great. With the current version from this branch and with all previous workarounds removed, all of our test cases pass. 🎉

Performance is also improved, which I could see in our two dedicated performance tests. These tests don't strictly measure pythonnet performance, but that is a significant part of them. Numbers are normalized to fractions, lower is better.

Variant	Test 1	Test 2
Python 3.12, pythonnet 3.0.5 + our own adaptation of filmors old DynamicObject workaround	1.0	1.0
Python 3.14, pythonnet 3.1.0rc0 + our own (problematic) workaround using codecs	1.18	1.04
Python 3.14, pythonnet with DLR support (this branch)	0.77	0.96

There is one significant behavioral difference that I observed: With past solutions, there was always an AttributeError when setting an attribute for which TrySetMember returns false (= which does not exist). With this branch, the attribute is now set on the Python side without any error. This can be problematic, because it can mask errors / typos in scripts and make troubleshooting harder. I don't know if this was intended. If it is, we can probably find a different solution.

I tried to find a workaround for that behavior and discovered something else where I'm not sure if it's correct: When throwing an exception in TrySetMember, the process crashes with a .NET exception, so there is no Python traceback to pinpoint the error location. In contrast, when throwing an exception in TryGetMember, it comes up as an AttributeError with Python traceback output.

Because we support running in different environments and Python versions, we probably need to implement automatic switching between DynamicObject workaround and first-class DLR support in pythonnet. Is there a way to find out if the currently loaded pythonnet module has DLR support? An information about the version would also work, but pythonnet currently has no __version__ attribute.

filmor · 2026-05-07T08:56:12Z

I'll try to handle these cases, definitely not intentional :)

If you need to get a package's version in your code, please use importlib.metadata.version().

@lostmsu

The dynamic getter swallowed any exception from TryGetMember and returned default to Python with the prior AttributeError still set, so user code observed a misleading AttributeError instead of the real failure. Set a Python exception in the catch arm. We use RuntimeError with the message string rather than Converter.ToPython(e) because wrapping the CLR exception object can trigger type initialisation that re-enters this same slot on the live dynamic object, producing infinite recursion. Mirrors the symmetry already present in the setter (pythonnet#2706 review, @lostmsu) and adds a regression test alongside the existing ThrowingSetDynamicObject coverage. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

@lostmsu

The dynamic getter swallowed any exception from TryGetMember and returned default to Python with the prior AttributeError still set, so user code observed a misleading AttributeError instead of the real failure. Set a Python exception in the catch arm. We use RuntimeError with the message string rather than Converter.ToPython(e) because wrapping the CLR exception object can trigger type initialisation that re-enters this same slot on the live dynamic object, producing infinite recursion. Mirrors the symmetry already present in the setter (pythonnet#2706 review, @lostmsu) and adds a regression test alongside the existing ThrowingSetDynamicObject coverage.

greateggsgreg · 2026-05-10T01:50:20Z

#2718 should fix the TryGetMember issue highlighted above.

@lostmsu

The dynamic getter swallowed any exception from TryGetMember and returned default to Python with the prior AttributeError still set, so user code observed a misleading AttributeError instead of the real failure. Set a Python exception in the catch arm. We use RuntimeError with the message string rather than Converter.ToPython(e) because wrapping the CLR exception object can trigger type initialisation that re-enters this same slot on the live dynamic object, producing infinite recursion. Mirrors the symmetry already present in the setter (#2706 review, @lostmsu) and adds a regression test alongside the existing ThrowingSetDynamicObject coverage.

alexchandel · 2026-05-11T15:00:30Z

You want to rebase this and retest with the TryGetMember fix?

- Cache HasClrMember reflection per (Type, name) so tp_getattro_dlr_proxy / tp_setattro_dlr_proxy avoid repeated GetMember() calls on every attribute access of DLR-aware objects. - Mirror tp_setattro_dlr_proxy's catch arm to the getter's safer SetError(RuntimeError, e.Message) shape instead of SetError(Exception), keeping both slots re-entry-safe on live dynamic objects. Related to pythonnet#2706.

@lostmsu

The dynamic getter swallowed any exception from TryGetMember and returned default to Python with the prior AttributeError still set, so user code observed a misleading AttributeError instead of the real failure. Set a Python exception in the catch arm. We use RuntimeError with the message string rather than Converter.ToPython(e) because wrapping the CLR exception object can trigger type initialisation that re-enters this same slot on the live dynamic object, producing infinite recursion. Mirrors the symmetry already present in the setter (#2706 review, @lostmsu) and adds a regression test alongside the existing ThrowingSetDynamicObject coverage.

- Cache HasClrMember reflection per (Type, name) so tp_getattro_dlr_proxy / tp_setattro_dlr_proxy avoid repeated GetMember() calls on every attribute access of DLR-aware objects. - Mirror tp_setattro_dlr_proxy's catch arm to the getter's safer SetError(RuntimeError, e.Message) shape instead of SetError(Exception), keeping both slots re-entry-safe on live dynamic objects. Related to #2706.

@lostmsu

The dynamic getter swallowed any exception from TryGetMember and returned default to Python with the prior AttributeError still set, so user code observed a misleading AttributeError instead of the real failure. Set a Python exception in the catch arm. We use RuntimeError with the message string rather than Converter.ToPython(e) because wrapping the CLR exception object can trigger type initialisation that re-enters this same slot on the live dynamic object, producing infinite recursion. Mirrors the symmetry already present in the setter (#2706 review, @lostmsu) and adds a regression test alongside the existing ThrowingSetDynamicObject coverage.

- Cache HasClrMember reflection per (Type, name) so tp_getattro_dlr_proxy / tp_setattro_dlr_proxy avoid repeated GetMember() calls on every attribute access of DLR-aware objects. - Mirror tp_setattro_dlr_proxy's catch arm to the getter's safer SetError(RuntimeError, e.Message) shape instead of SetError(Exception), keeping both slots re-entry-safe on live dynamic objects. Related to #2706.

@lostmsu

The dynamic getter swallowed any exception from TryGetMember and returned default to Python with the prior AttributeError still set, so user code observed a misleading AttributeError instead of the real failure. Set a Python exception in the catch arm. We use RuntimeError with the message string rather than Converter.ToPython(e) because wrapping the CLR exception object can trigger type initialisation that re-enters this same slot on the live dynamic object, producing infinite recursion. Mirrors the symmetry already present in the setter (#2706 review, @lostmsu) and adds a regression test alongside the existing ThrowingSetDynamicObject coverage.

- Cache HasClrMember reflection per (Type, name) so tp_getattro_dlr_proxy / tp_setattro_dlr_proxy avoid repeated GetMember() calls on every attribute access of DLR-aware objects. - Mirror tp_setattro_dlr_proxy's catch arm to the getter's safer SetError(RuntimeError, e.Message) shape instead of SetError(Exception), keeping both slots re-entry-safe on live dynamic objects. Related to #2706.

filmor · 2026-05-13T10:28:41Z

@spkl Can you please test out the latest version?
@greateggsgreg I stole your pending commit that includes the caching, I hope that's fine :)
@lostmsu One last round of review? Once this is through, I would merge this and publish it as 3.1.0-rc2, we can discuss on a separate issue (after I update the Changelog to reflect the actual changes) whether we want to release it as 4.0. I don't want to delay the release any further :)

spkl · 2026-05-13T15:35:56Z

@filmor All good from my side 👍

- Catch exceptions in TrySet/DeleteMember - Convert the exceptions into Python exceptions - Add tests for the remaining cases - Add a note on why the field has to be lazily initialized (general issue with derived classes)

@lostmsu

The dynamic getter swallowed any exception from TryGetMember and returned default to Python with the prior AttributeError still set, so user code observed a misleading AttributeError instead of the real failure. Set a Python exception in the catch arm. We use RuntimeError with the message string rather than Converter.ToPython(e) because wrapping the CLR exception object can trigger type initialisation that re-enters this same slot on the live dynamic object, producing infinite recursion. Mirrors the symmetry already present in the setter (#2706 review, @lostmsu) and adds a regression test alongside the existing ThrowingSetDynamicObject coverage.

- Cache HasClrMember reflection per (Type, name) so tp_getattro_dlr_proxy / tp_setattro_dlr_proxy avoid repeated GetMember() calls on every attribute access of DLR-aware objects. - Mirror tp_setattro_dlr_proxy's catch arm to the getter's safer SetError(RuntimeError, e.Message) shape instead of SetError(Exception), keeping both slots re-entry-safe on live dynamic objects. Related to #2706.

filmor requested a review from lostmsu April 18, 2026 13:51

filmor force-pushed the dlr branch 3 times, most recently from 34eff30 to ac40869 Compare April 18, 2026 17:58

filmor marked this pull request as ready for review April 20, 2026 07:23

lostmsu reviewed Apr 21, 2026

View reviewed changes

Comment thread src/runtime/Types/ClassBase.cs Outdated

Comment thread src/runtime/Types/ClassBase.cs Outdated

Comment thread src/runtime/Types/ClassBase.cs Outdated

Comment thread src/runtime/Types/ClassBase.cs Outdated

Comment thread src/runtime/Types/ClassBase.cs Outdated

filmor force-pushed the dlr branch 4 times, most recently from fdfa219 to 6d1572e Compare May 5, 2026 19:56

filmor force-pushed the dlr branch from 6d1572e to 88e8bd3 Compare May 6, 2026 13:43

filmor force-pushed the dlr branch from 88e8bd3 to d629b87 Compare May 7, 2026 17:01

filmor requested a review from lostmsu May 7, 2026 17:02

greateggsgreg mentioned this pull request May 9, 2026

dlr branch fix - Propagate exceptions from TryGetMember in tp_getattro_dlr_proxy #2718

Merged

filmor force-pushed the dlr branch from da48ef7 to 93c8e38 Compare May 12, 2026 19:56

filmor force-pushed the dlr branch from fef6f35 to e99538e Compare May 12, 2026 20:25

filmor force-pushed the dlr branch from e99538e to a3c2069 Compare May 13, 2026 05:28

filmor and others added 5 commits May 15, 2026 21:57

Implement support for DLR get/set

5dacfb4

Use explicit IsNull on BorrowedReference

69ba861

Catch errors in setting/deleting properties

2408c43

- Catch exceptions in TrySet/DeleteMember - Convert the exceptions into Python exceptions - Add tests for the remaining cases - Add a note on why the field has to be lazily initialized (general issue with derived classes)

filmor force-pushed the dlr branch from a3c2069 to 3a08990 Compare May 15, 2026 19:57

filmor merged commit 27caf5c into master May 15, 2026
29 of 30 checks passed

filmor deleted the dlr branch May 15, 2026 19:58

Conversation

filmor commented Apr 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

spkl commented Apr 20, 2026

Uh oh!

filmor commented Apr 20, 2026

Uh oh!

spkl commented Apr 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

filmor commented May 6, 2026

Uh oh!

spkl commented May 6, 2026

Uh oh!

spkl commented May 7, 2026

Uh oh!

filmor commented May 7, 2026

Uh oh!

greateggsgreg commented May 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

alexchandel commented May 11, 2026

Uh oh!

filmor commented May 13, 2026

Uh oh!

spkl commented May 13, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

filmor commented Apr 18, 2026 •

edited

Loading

spkl commented Apr 20, 2026 •

edited

Loading

greateggsgreg commented May 10, 2026 •

edited

Loading