Our journey continues through our detailed Python Exception Handling series with a deep look at the NameError found in Python. Just as with many other programming languages, Python source code (typically found in .py
files) is initially compiled
into bytecode
, which is a low level representation of source code that can be executed by a virtual machine via the CPython interpreter. Part of this process involves loading local
or global
objects into the callstack. However, when Python attempts to load an object that doesn't exist elsewhere in the callstack it will forcefully raise a NameError
indicating as much.
In today's article we'll explore the NameError
by looking at where it resides in the overall Python Exception Class Hierarchy. We'll also look at some functional sample code that illustrates the basic compilation process Python source code goes through to turn into bytecode
, and how improper references can result in NameErrors
during this process. Let's get right into it!
The Technical Rundown
All Python exceptions inherit from the BaseException
class, or extend from an inherited class therein. The full exception hierarchy of this error is:
BaseException
Exception
NameError
Full Code Sample
Below is the full code sample we'll be using in this article. It can be copied and pasted if you'd like to play with the code yourself and see how everything works.
import datetime
import disfrom gw_utility.book import Book
from gw_utility.logging import Loggingdef main():
try:
# Create Book.
book = Book("The Hobbit", "J.R.R. Tolkien", 366, datetime.date(1937, 9, 15))# Log book object.
Logging.line_separator("log_object(book)", 60)
log_object(book)# Log invalid object.
Logging.line_separator("log_invalid_object(book)", 60)
log_invalid_object(book)# Disassemble both log_ functions.
Logging.line_separator("DISASSEMBLY OF log_object()", 60)
disassemble_object(log_object)Logging.line_separator("DISASSEMBLY OF log_invalid_object()", 60)
disassemble_object(log_invalid_object)
except NameError as error:
# Output expected NameErrors.
Logging.log_exception(error)
except Exception as exception:
# Output unexpected Exceptions.
Logging.log_exception(exception, False)def log_object(value):
"""Logs passed value parameter to console.:param value: Value to be logged.
:return: None
"""
try:
Logging.log(value)
except NameError as error:
# Output expected NameErrors.
Logging.log_exception(error)
except Exception as exception:
# Output unexpected Exceptions.
Logging.log_exception(exception, False)def log_invalid_object(value):
"""Attempts to log invalid object (valu) to console.:param value: Value intended to be logged, but which is instead ignored.
:return: None
"""
try:
Logging.log(valu)
except NameError as error:
# Output expected NameErrors.
Logging.log_exception(error)
except Exception as exception:
# Output unexpected Exceptions.
Logging.log_exception(exception, False)def disassemble_object(value):
"""Outputs disassembly of passed object.:param value: Object to be disassembled.
:return: None
"""
dis.dis(value)if __name__ == "__main__":
main()
# book.py
import datetimeclass Book:
author: str
page_count: int
publication_date: datetime.date
title: strdef __eq__(self, other):
"""Determines if passed object is equivalent to current object."""
return self.__dict__ == other.__dict__def __init__(self,
title: str = None,
author: str = None,
page_count: int = None,
publication_date: datetime.date = None):
"""Initializes Book instance.:param title: Title of Book.
:param author: Author of Book.
:param page_count: Page Count of Book.
:param publication_date: Publication Date of Book.
"""
self.author = author
self.page_count = page_count
self.publication_date = publication_date
self.title = titledef __getattr__(self, name: str):
"""Returns the attribute matching passed name."""
# Get internal dict value matching name.
value = self.__dict__.get(name)
if not value:
# Raise AttributeError if attribute value not found.
raise AttributeError(f'{self.__class__.__name__}.{name} is invalid.')
# Return attribute value.
return valuedef __len__(self):
"""Returns the length of title."""
return len(self.title)
def __str__(self):
"""Returns a formatted string representation of Book."""
date = '' if self.publication_date is None else f', published on {self.publication_date.__format__("%B %d, %Y")}'
pages = '' if self.page_count is None else f' at {self.page_count} pages'
return f'\'{self.title}\' by {self.author}{pages}{date}.'
This code sample also uses the Logging
utility class, the source of which can be found here on GitHub.
When Should You Use It?
As mentioned in the introduction, a NameError
will occur when the CPython interpreter does not recognize a local
or global
object name that has been provided in the Python source code. Let's jump right into some example code in normal Python, after which we'll see how we can disassemble this code into the bytecode
that CPython actually reads and interprets.
We begin with two extremely simple functions, log_object(value)
and log_invalid_object(value)
:
def log_object(value):
"""Logs passed value parameter to console.:param value: Value to be logged.
:return: None
"""
try:
Logging.log(value)
except NameError as error:
# Output expected NameErrors.
Logging.log_exception(error)
except Exception as exception:
# Output unexpected Exceptions.
Logging.log_exception(exception, False)def log_invalid_object(value):
"""Attempts to log invalid object (valu) to console.
:param value: Value intended to be logged, but which is instead ignored.
:return: None
"""
try:
Logging.log(valu)
except NameError as error:
# Output expected NameErrors.
Logging.log_exception(error)
except Exception as exception:
# Output unexpected Exceptions.
Logging.log_exception(exception, False)
The majority of the code for each of these is merely there for error handling, as the core functionality takes place on a single line: Logging.log(value)
and Logging.log(valu)
, respectively. In essence, we're merely using these two functions to log the content of the passed value
parameter to the console. However, in the case of log_invalid_object()
we have a slight typo of valu
instead of value
.
Let's test these out by creating a simple Book
object instance and passing it to each of our two log_
functions:
# Create Book.
book = Book("The Hobbit", "J.R.R. Tolkien", 366, datetime.date(1937, 9, 15))# Log book object.
Logging.line_separator("log_object(book)", 60)
log_object(book)
# Log invalid object.
Logging.line_separator("log_invalid_object(book)", 60)
log_invalid_object(book)
As you can probably guess, executing this code produces an expected Book
object output, followed by raising a NameError
because our typo of valu
is not a recognized name:
--------------------- log_object(book) ---------------------
'The Hobbit' by J.R.R. Tolkien at 366 pages, published on September 15, 1937.
----------------- log_invalid_object(book) -----------------
[EXPECTED] NameError: name 'valu' is not defined
That's all well and good, but Python is a powerful language that allows us to look "under the hood" a bit and see the actual bytecode
that each of these log_
functions generates for the CPython interpreter. We'll be using the built-in dis
disassembler module, which was created for this very purpose. By passing a function reference to the dis.dis()
method we are provided a full output of the disassembled bytecode
that CPython interprets during execution. Our local disassemble_object(value)
function is a small wrapper for this purpose:
def disassemble_object(value):
"""Outputs disassembly of passed object.
:param value: Object to be disassembled.
:return: None
"""
dis.dis(value)
Thus, we can see what the bytecode
of the log_object(value)
function looks like by running the following:
# Disassemble both log_ functions.
Logging.line_separator("DISASSEMBLY OF log_object()", 60)
disassemble_object(log_object)
This produces the following output:
--------------- DISASSEMBLY OF log_object() ----------------
41 0 SETUP_EXCEPT 14 (to 16)42 2 LOAD_GLOBAL 0 (Logging)
4 LOAD_ATTR 1 (log)
6 LOAD_FAST 0 (value)
8 CALL_FUNCTION 1
10 POP_TOP
12 POP_BLOCK
14 JUMP_FORWARD 88 (to 104)43 >> 16 DUP_TOP
18 LOAD_GLOBAL 2 (NameError)
20 COMPARE_OP 10 (exception match)
22 POP_JUMP_IF_FALSE 58
24 POP_TOP
26 STORE_FAST 1 (error)
28 POP_TOP
30 SETUP_FINALLY 16 (to 48)45 32 LOAD_GLOBAL 0 (Logging)
34 LOAD_ATTR 3 (log_exception)
36 LOAD_FAST 1 (error)
38 CALL_FUNCTION 1
40 POP_TOP
42 POP_BLOCK
44 POP_EXCEPT
46 LOAD_CONST 1 (None)
>> 48 LOAD_CONST 1 (None)
50 STORE_FAST 1 (error)
52 DELETE_FAST 1 (error)
54 END_FINALLY
56 JUMP_FORWARD 46 (to 104)46 >> 58 DUP_TOP
60 LOAD_GLOBAL 4 (Exception)
62 COMPARE_OP 10 (exception match)
64 POP_JUMP_IF_FALSE 102
66 POP_TOP
68 STORE_FAST 2 (exception)
70 POP_TOP
72 SETUP_FINALLY 18 (to 92)
48 74 LOAD_GLOBAL 0 (Logging)
76 LOAD_ATTR 3 (log_exception)
78 LOAD_FAST 2 (exception)
80 LOAD_CONST 2 (False)
82 CALL_FUNCTION 2
84 POP_TOP
86 POP_BLOCK
88 POP_EXCEPT
90 LOAD_CONST 1 (None)
>> 92 LOAD_CONST 1 (None)
94 STORE_FAST 2 (exception)
96 DELETE_FAST 2 (exception)
98 END_FINALLY
100 JUMP_FORWARD 2 (to 104)
>> 102 END_FINALLY
>> 104 LOAD_CONST 1 (None)
106 RETURN_VALUE
This may appear a bit overwhelming at first, but this data is actually quite easy to interpret with a bit of knowledge about what we're looking at in each column. The first column (e.g. 41
, 42
, 43
... 48
) is the actual line number in the source doe for the corresponding set of instructions. Thus, we can see that all of the following instructions...
42 2 LOAD_GLOBAL 0 (Logging)
4 LOAD_ATTR 1 (log)
6 LOAD_FAST 0 (value)
8 CALL_FUNCTION 1
10 POP_TOP
12 POP_BLOCK
14 JUMP_FORWARD 88 (to 104)
...were generated from a single line of source code (#42):
Logging.log(value)
The column with multiples of two (0
, 2
, 4
, etc) is the memory address
in the underlying bytecode
for the given instruction. Modern Python stores instructions using two bytes of data, hence the multiples of two. The next column contains the opname
(i.e. instruction) that should be executed, all of which can be found in the official documentation.
The column after that contains any arguments
, if applicable, that each particular instruction will use. The final column provides a human-friendly version of the instruction, so we can better visualize how the bytecode
instruction correlates to source code.
Thus, let's look back at the single line 42
source code of Logging.log(value)
and the generated bytecode
instruction set to see what's going on:
42 2 LOAD_GLOBAL 0 (Logging)
4 LOAD_ATTR 1 (log)
6 LOAD_FAST 0 (value)
8 CALL_FUNCTION 1
10 POP_TOP
12 POP_BLOCK
14 JUMP_FORWARD 88 (to 104)
It starts with LOAD_GLOBAL
to load the global name Logging
onto the stack. It then loads the log
attribute onto the top of the stack (TOS
). LOAD_FAST
pushes a reference to a local
variable called value
onto the stack. Next, CALL_FUNCTION
calls the function at argument stack 1
, which is the log
method added two instructions prior. POP_TOP
removes the most recent item added onto the stack, which is the local
value
object. Every frame of execution contains a stack of code blocks
, which are the logical groupings we see and create when writing source code that is locally grouped. For example, a nested loop
or, in this case, a try-except
block, is contained within a separate code block in the stack. Since the next instruction that we're jumping to with JUMP_FORWARD 88
is exiting the end of the try
block found in our source code, POP_BLOCK
is used to remove the top (current) block from the code block stack.
Cool, so let's see how this compiled bytecode
for log_object
differs from the slightly modified log_invalid_object
function:
Logging.line_separator("DISASSEMBLY OF log_invalid_object()", 60)
disassemble_object(log_invalid_object)
We'll ignore the majority of the bytecode
produced here since it is identical to that produced by log_object
, but here we have the instruction set from the same corresponding Logging.log(valu)
source code line we examined before:
58 2 LOAD_GLOBAL 0 (Logging)
4 LOAD_ATTR 1 (log)
6 LOAD_GLOBAL 2 (valu)
8 CALL_FUNCTION 1
10 POP_TOP
12 POP_BLOCK
14 JUMP_FORWARD 88 (to 104)
Everything looks exactly the same as before with two exceptions: The line number
of 58
is obviously different, since we're compiling a different line of source code. The second difference is the third instruction, which changed from LOAD_FAST 0 (value)
to LOAD_GLOBAL 2 (valu)
. Why? Because the compiler cannot reconcile a local
object named valu
, since the actual local
parameter passed into the function is value
, without the typo. Therefore, the compiler assumes valu
is a global
name, and tries to load it via LOAD_GLOBAL
. As we know from executing the log_invalid_object
function earlier, the CPython interpreter is unable to locate a global
named valu
during execution, so a NameError
is raised to indicate as much. Neat!
Airbrake's robust error monitoring software provides real-time error monitoring and automatic exception reporting for all your development projects. Airbrake's state of the art web dashboard ensures you receive round-the-clock status updates on your application's health and error rates. No matter what you're working on, Airbrake easily integrates with all the most popular languages and frameworks. Plus, Airbrake makes it easy to customize exception parameters, while giving you complete control of the active error filter system, so you only gather the errors that matter most.
Check out Airbrake's error monitoring software today with a free 14-day trial, and see for yourself why so many of the world's best engineering teams use Airbrake to revolutionize their exception handling practices!