Repository navigation
Offer suggestions on AttributeError and NameError #82711
Description
Activity
To improve the debugging experience in both interactive and non-interactive code, I propose to offer suggestions when attribute access fails. For example:
>>> class A: foo = None ... >>> A.fou Traceback (most recent call last): File "<stdin>", line 1, in <module> AttributeError: type object 'A' has no attribute 'fou'
Did you mean: foo?
This also applies to imports from modules and other situations:
>>> import collections >>> collections.NamedTuple Traceback (most recent call last): File "<stdin>", line 1, in <module> AttributeError: module 'collections' has no attribute 'NamedTuple'Did you mean: namedtuple?
It already exists as a 3rd party module and it would be really cool to have this in core level.
https://github.com/dutc/didyoumean (by James Powell)
I am not super convinced that this is a great idea because it has some performance cost (although somewhat controlled) but I want to open a discussion.
Ruby has it integrated into the core : https://bugs.ruby-lang.org/issues/11252 . It was initially a gem that got merged into core.
methosd
'
undefined local variable or methodmethosd' for main:Object Did you mean? methods method (repl):1:inIdea: we could only do this in interactive mode if we consider that is expensive enough.
I am running pyperformance to check the performance cost of this.
Slower (27):
- pathlib: 25.2 ms +- 0.4 ms -> 105 ms +- 2 ms: 4.18x slower (+318%)
- sympy_str: 315 ms +- 3 ms -> 500 ms +- 3 ms: 1.59x slower (+59%)
- sympy_sum: 203 ms +- 2 ms -> 286 ms +- 2 ms: 1.41x slower (+41%)
- sqlalchemy_imperative: 36.5 ms +- 0.9 ms -> 41.8 ms +- 0.8 ms: 1.14x slower (+14%)
- python_startup: 12.4 ms +- 0.1 ms -> 13.9 ms +- 0.0 ms: 1.12x slower (+12%)
- python_startup_no_site: 9.19 ms +- 0.04 ms -> 10.3 ms +- 0.0 ms: 1.12x slower (+12%)
- sympy_integrate: 24.9 ms +- 0.1 ms -> 27.6 ms +- 0.3 ms: 1.11x slower (+11%)
- scimark_sparse_mat_mult: 5.21 ms +- 0.05 ms -> 5.49 ms +- 0.06 ms: 1.05x slower (+5%)
- unpickle_list: 5.78 us +- 0.07 us -> 6.08 us +- 0.07 us: 1.05x slower (+5%)
- 2to3: 392 ms +- 2 ms -> 411 ms +- 10 ms: 1.05x slower (+5%)
- nbody: 155 ms +- 2 ms -> 160 ms +- 2 ms: 1.03x slower (+3%)
- scimark_fft: 423 ms +- 2 ms -> 432 ms +- 3 ms: 1.02x slower (+2%)
- float: 142 ms +- 2 ms -> 145 ms +- 2 ms: 1.02x slower (+2%)
- unpack_sequence: 68.6 ns +- 2.5 ns -> 69.9 ns +- 2.1 ns: 1.02x slower (+2%)
- pickle_list: 5.65 us +- 0.05 us -> 5.76 us +- 0.07 us: 1.02x slower (+2%)
- fannkuch: 569 ms +- 3 ms -> 579 ms +- 6 ms: 1.02x slower (+2%)
- xml_etree_parse: 196 ms +- 2 ms -> 199 ms +- 2 ms: 1.02x slower (+2%)
- sqlalchemy_declarative: 210 ms +- 2 ms -> 213 ms +- 3 ms: 1.01x slower (+1%)
- unpickle: 16.9 us +- 0.1 us -> 17.2 us +- 0.7 us: 1.01x slower (+1%)
- regex_effbot: 3.93 ms +- 0.12 ms -> 3.97 ms +- 0.04 ms: 1.01x slower (+1%)
- scimark_monte_carlo: 128 ms +- 2 ms -> 129 ms +- 2 ms: 1.01x slower (+1%)
- scimark_lu: 186 ms +- 5 ms -> 188 ms +- 6 ms: 1.01x slower (+1%)
- regex_dna: 272 ms +- 1 ms -> 274 ms +- 1 ms: 1.01x slower (+1%)
- xml_etree_iterparse: 130 ms +- 2 ms -> 131 ms +- 3 ms: 1.01x slower (+1%)
- genshi_xml: 74.3 ms +- 0.8 ms -> 74.8 ms +- 0.8 ms: 1.01x slower (+1%)
- regex_v8: 29.7 ms +- 0.3 ms -> 29.9 ms +- 0.2 ms: 1.01x slower (+1%)
- mako: 19.3 ms +- 0.3 ms -> 19.4 ms +- 0.2 ms: 1.01x slower (+1%)
The current approach is too expensive, so I'm closing PR 16850.
AFAIK there is existing issue for this idea.
I have doubts about performance. I added _PyObject_LookupAttr in particularly to avoid an overhead of raising and silencing an AttributeError. I believe most performance sensitive code in the core now uses it and will not be affected, but there is other code which silences it (_PyObject_LookupAttr itself silences an AttributeError raised in called functions), third-party code which uses PyObject_HasAttr or PyObject_GetAttr can be affected.
It might be simpler to implement it in sys.excepthook or sys.displayhook, but at that point we do not have attribute name and a reference to the object. There is an issues and a PEP about adding references to AttributeError. It could help to implement this feature.
I am surprised that it was SO expensive.
Pathlib would largely benefit from cached_property if it be compatible with slots.
Serhiy, do you think we could attach the object and the name to some private fields of the AttributeError and check that in sys.excepthook if they are present?
I will also repeat the pyperformance results locally just in case something was off on the speed.python.org server.
17 remaining items
I like the https://docs.python.org/dev/whatsnew/3.10.html#better-error-messages section: well done, thanks ;-)
I opened PR 25584 to fix this current behavior:
>>> v Traceback (most recent call last): File "<stdin>", line 1, in <module> NameError: name 'v' is not defined. Did you mean: 'id'? >>> vv Traceback (most recent call last): File "<stdin>", line 1, in <module> NameError: name 'vv' is not defined. Did you mean: 'id'? >>> vvv Traceback (most recent call last): File "<stdin>", line 1, in <module> NameError: name 'vvv' is not defined. Did you mean: 'abs'?
Some research of other projects:
LLVM [1][2]
------------ Compute Levenshtein
- Using O(n) memory rather than O(n^2)
- Uses UpperBound = (len(typo) + 2) // 3
GCC [3]
-------- Uses Damerau-Levenshtein distance
- Counts transpositions like "abcd" <-> "bacd" as one move
- Swapping Case as in "a" <-> "A" counts as half a move
- cutoff = (longer + 2) // 3 if longer - shorter >= 2 else max(longer // 3, 1)
Rust [4]
--------- "maximum allowable edit distance defaults to one-third of the given word."
- First checks for exact case-insensitive match, then check for Levenshtein distance small enough, then check if sorted(a.split("_")) == sorted(b.split("_"))
Ruby [5]
--------- Quickly filter out words with bad Jaro–Winkler distance
- threshold = input.length > 3 ? 0.834 : 0.77
- Only compute Levenshtein for words that remain
- threshold = (input.length * 0.25).ceil
- Output all good enough words
- If no word was good enough then output the closest match.
I think there are some good ideas here.
[1] https://github.com/llvm/llvm-project/blob/d480f968ad8b56d3ee4a6b6df5532d485b0ad01e/llvm/include/llvm/ADT/edit_distance.h#L42
[2] https://github.com/llvm/llvm-project/blob/e2b3b89bf1ce74bf889897e0353a3e3fa93e4452/clang/lib/Sema/SemaLookup.cpp#L4263
[3] https://github.com/gcc-mirror/gcc/blob/16e2427f50c208dfe07d07f18009969502c25dc8/gcc/spellcheck.c
[4] https://github.com/rust-lang/rust/blob/673d0db5e393e9c64897005b470bfeb6d5aec61b/compiler/rustc_span/src/lev_distance.rs#L44
[5] https://github.com/ruby/ruby/blob/48b94b791997881929c739c64f95ac30f3fd0bb9/lib/did_you_mean/spell_checker.rb- Compute Levenshtein
Hi Dennis, this is a fantastic investigation!
I think I really like GCC approach here. We may want to invest into porting some of their ideas into our solution.
PR 25776 is a work in progress for what it might look like to do a few things:
- Make case-swaps half the cost of any other edit
- Refactor Levenshtein code to not use memory allocator, and to bail early on no match.
- Add comments to Levenshtein distance code
- Add test cases for Levenshtein distance behind a debug macro
- Set threshold to (name_size + item_size + 3) * MOVE_COST / 6.
- Reasoning: similar to difflib.SequenceMatcher.ratio() >= 2/3:
"Multiset Jaccard similarity" >= 2/3
matching letters / total letters >= 2/3
(name_size - distance + item_size - distance) / (name_size + item_size) >= 2/3
1 - (2*distance) / (name_size + item_size) >= 2/3
1/3 >= (2*distance) / (name_size + item_size)
(name_size + item_size) / 6 >= distance
With rounding:
(name_size + item_size + 3) // 6 >= distance
- Reasoning: similar to difflib.SequenceMatcher.ratio() >= 2/3:
Re: Damerau-Levenshtein (transpositions as single edits), if that were to get implemented, I don't see a way to do that without using a buffer of at least 3x the size, storing the most recent 3 rows of the matrix.
Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.
Show more details
GitHub fields:
bugs.python.org fields: