Class Statement vs Dynamic Class Creation

We know that along with the standard class statement, Python also allows us to create classes dynamically by calling type() (or another metaclass if our class has a metaclass other than type)

I already dedicated a rather thick post to type. Basically we use it for creating a new class like this: type(classname, superclasses, namespace). (the namespace is just a dictionary with the attributes).

So I was wondering if the compiler translates a class statement into a call to type(), and yes, more or less we can say so, but there are some extras. I've had a really insightful conversation with a GPT about this, and additionally I've found an excellent article that explains it in full detail

The steps that Python follows when it comes across with a class statement (class Foo(Base, metaclass=Meta): x = 1) are these (I'm taking it from a GPT discussion, it's basically the same that is explained in the linked article)

• Step 1 — Determine the metaclass. Python calls __build_class__ (a builtin), which inspects the bases and the explicit metaclass= kwarg to resolve which metaclass to use (with MRO-based metaclass conflict resolution).
• Step 2 — Prepare the namespace. The metaclass's __prepare__ classmethod is called: namespace = Meta.__prepare__('Foo', (Base,), **kwargs). This returns the dict (or dict-like object) that will serve as the class namespace. For type, this is just a plain dict. For enum.EnumMeta, for example, it returns a special _EnumDict.
• Step 3 — Execute the class body. The compiled code object for the class body is executed as a function, with the namespace from step 2 as its locals(). This is the key insight: it's essentially exec(body_code, globals(), namespace). After this, namespace contains {'x': 1, '__module__': ..., '__qualname__': ...}.
• Step 4 — Call the metaclass. Meta('Foo', (Base,), namespace) is called — which, for the default type, invokes type.__call__ → type.__new__ → type.__init__. This is where the actual class object is constructed.

How do the above steps look at the bytecode level? When the Python compiler (yes, in Python, where compilation is like a hidden step that happens the first time our code is run (or has changed), it's sometimes confusing to establish the difference between compilation time and execution time), comes across a class statement, it creates a code object for the code that we've placed inside that statement (the body of the class statement), along with code objects for each function (method) defined in that code, and creates a sequence of bytecode instructions that at runtime will make use of that code object (and many more things) to create a class object (yes, remember that classes are objects).

That sequence of bytecode instructions can vary slightly with Python versions (what I'll show below, that corresponds to python 3.14 is slightly different from what is shown in the aforementioned article), but the intent is the same.

class Person:
	pass

# translates into:

 0           RESUME                   0

  1           LOAD_BUILD_CLASS
              PUSH_NULL
              LOAD_CONST               0 (code object Person at 0x78d07576e730, file "class_creation.py", line 1)
              MAKE_FUNCTION
              LOAD_CONST               1 ('Person')
              CALL                     2
              STORE_NAME               0 (Person)
              LOAD_CONST               2 (None)
              RETURN_VALUE

So those seem like very few instructions for the complex 4 steps that I've just described!. Well, that's because all the magic happens in a builtin function __build_class, that is loaded by the LOAD_BUILD_CLASS bytecode instruction. The article makes a great job explaining these opcodes.

When we create a class dynamically using type() (or any other metaclass), we are directly at step 4, we are skipping the first 3 steps. Obviously it's us who choose the metaclass to use, and there's not class body to execute. And as we create ourselves the namespace object to pass to the metaclass the __prepare__ method that helps prepare that namespace is not executed. That's maybe the main difference then, that in the dynamic class creation the metaclass __prepare__ method does not intervene. That's interesting, cause indeed I was not familiar with that __prepare__ method (also referred as hook).

When talking about metaclasses I always think about __new__ and __init__ (and __call__ that intervenes when instances of a class created by the metaclass are created), I've talked about them in different posts, one of the most interesting being this, but was unfamiliar with __prepare__. We've seen that it allows us to prepare the namespace, OK, but when can we need that? Well, very rarely (this is particularly dark metaclass stuff). That can be material for another post, for now I'll just say that enum.EnumMeta makes use of it.

SQL, NULL, Unknown

Lately I've been revisiting the rather particular behaviour of NULL in SQL, and it has led me into a better understanding of how different SQL is from General Programming Languages

- Binary Logic vs Ternary Logic

General Programming Languages (Python, JavaScript, ruby, Java...) use binary logic (Boolean logic in particular, and indeed that's the only logic I was aware of). Conditions are either True or False.
SQL uses a ternary logic (kleene logic), where we have TRUE, FALSE, and UNKNOWN

- The meaning of Missing Data

Both in General Programming Languages and in SQL we use null (None in Python) to represent missing data. There are 2 reasons for missing data, either it does not apply to that object, or we don't know it. Let's say we have an instance of a ShopItem class. Its expirationDate attribute can be null either because this object is a Book, and books do not expire, or because the printed date on this beans can is blurry (or we've had no time to read it yet) and then we don't know it, it's unknown.

In general Programming Languages null is a value (that represents that there's nothing here, there is no value here, for whatever the reason, either because it does not apply or because we don't know it), and with binary logic comparing a value to to another value is either true or false. So "a" == null is false, and null == null is true.

In SQL we have a sort of mismatch. On one hand we have ternary logic with that additional UNKOWN concept, but on the other hand we still have a single value, NULL, to represent both that it does not apply or that we don't know it. So how should NULL behave in comparisons? SQL designers decided to treat NULL as a marker that represents that the value is unknown (so we can not express that the value does not apply).

Once we have understood that, the apparent odd behaviour of NULL in comparisons suddenly makes sense. Any comparison using the standard operators (=, !=, <, >, <>) involving a NULL value will return UNKNOWN, even NULL = NULL or NULL != NULL return UNKNOWN. The negation of UNKNOWN (NOT UNKNWON) is also UNKNOWN.

What is odd is what I've just said, that SQL lacks a way to indicate that the value does not apply. It seems one of the main influences in the design of SQL ended realizing this was a serious problem, but too late:

Codd actually realized this flaw later in his life and proposed that SQL should have two different kinds of NULLs: A-Values (Absence) and I-Values (Information Unknown). Sadly, by then, SQL was already set in stone.

Context Managers Part 1

Context Managers have existed in Python since version 2.5, while Assignment Expressions (walrus operator) were added in version 3.8. Somehow recently I came up to wondering if we can replace the "as" by a ":=" assignment. I mean, can we do this?:

with it := MyContextManager():
	# do whatever with it

rather than this:

with MyContextManager() as it:
	# do whatever with it

The answer is NO, or well, more accurately, sometimes yes, sometimes no, but you should always avoid it. To understand this we have to review what a Context Manager is and how they work. Notice that they're part of a broader concept: Automatic Resource Management that also includes Garbage Collecion and RAII (Resource Acquisition Is Initialization).

A Python Context Manager handles the setup and cleanup of resources in your programs. A context manager is any object that implements:

__enter__(self)
__exit__(self, exc_type, exc_value, traceback)

And it's used like this:

with EXPR as target:
    BODY

And now the important part. Conceptually, Python does something roughly like this (as explained by a GPT):

resource_manager = EXPR
resource = resource_manager.__enter__()
try:
    target = resource
    BODY
finally:
    manager.__exit__(...)

So the key point is: The object used to manage the context and the object bound after as do not have to be the same object. That is exactly why __enter__() is allowed to return anything.

Some Context Managers are implemented so that __enter__ returns the context manager itself, while others return a different object. Basically, in the first case the resource being managed and the Resource Manager (Context Manager) are the same object, in the second case they are different as the management responsability has been moved away from the resource itself, to a different object.

Another interesting topic. We know that in Python a single with block can include multiple context managers. I mean:

with open('a.txt', 'r') as fr, open('b.txt', 'w') as fw:
    do_something(fr, fw)

I was wondering if cases where the second context manager makes use of the first context manager, and that I think is more common to find written like this:

with ContextManager1("aaa") as ctx1:
	with ContextManager2(ctx1) as ctx2:
    		do_something(ctx1, ctx2)

Could be written with a single with (avoiding the additional nesting level):

with ContextManager1("aaa") as ctx1, ContextManager2(ctx1) as ctx2:
	do_something(ctx1, ctx2)

The answer is YES. The first

Python evaluates multiple context managers in a single with statement sequentially from left to right. The moment the first context manager is entered, its return value is bound to the as variable, making it immediately available for the next context manager on the same line.

Type Hints Notes 2026

Type hints are more and more prevalent in recent Python code. I'm still not too severe about them, but my level of strictness continues to grow over time. I've lately learnt a couple of things:

Variadic Parameters. When typing functions that have packed (variadic) parameters in its signature (*args, **kwargs), we put the type of the individual parameter, we don't have to type it as a collection or dictionary (save if each parameter is really a collection), I mean, for a pipe function we should do:

# this is RIGHT
def pipe(val: Any, *fns: Callable) -> Any:

# this is WRONG
def pipe(val: Any, *fns: list[Callable]) -> Any:

Tuples. In Python we use tuples for "groups" of a fixed number of elements, a pair, a trio... We express it in the signature like this: tuple[str, str] or tuple[int, str, str]... But how to express that a function returns (or receives) a "group" of an unknown number of elements? We also use tuples, combined with ellipsis (...), like this:

tuple[int, ...]         # any number of ints: (), (1,), (1, 2, 99), ...
tuple[int | str, ...]   # any number of elements, where each element can be an int or a str (), ("a", 1), ("a", "b", "c"), (1, 2, 1) ...
  
tuple[int, str, bool]   # exactly 3 elements: an int, a str, a bool
  

An important detail that I've learnt thanks to a typing issue. We know that in a Python try-except block, the except clause can manage multiple exception types, I mean: except RuntimeError, TypeError, NameError:. Those multiple exceptions are a tuple, not just any iterable. Let's see an example (the last line is what is WRONG):

# multiple exceptions
def multiple_exceptions(exceptions: tuple[type[Exception], ...]) -> None:
    try:
        raise ValueError("This is a ValueError")
    except exceptions as e:
        print(f"Caught an exception: {e}")

multiple_exceptions((ValueError, TypeError))  
# Caught an exception: ValueError

# important, this is WRONG, we have to pass a tuple, not just any collection
multiple_exceptions([ValueError, TypeError])
# TypeError: catching classes that do not inherit from BaseException is not allowed

Indeed, an equivalent function with a variadic signature feels more natural and idiomatic than the above (and furthermore prevents the confusion of passing over any collection rather than exactly a tuple):

# this variadic signature feels more natural
def multiple_exceptions2(*exceptions: type[Exception]) -> None:
    if not exceptions:
        raise ValueError("pass at least one exception type")
    try:
        raise ValueError("This is a ValueError")
    except exceptions as e:
        print(f"Caught an exception: {e}")

multiple_exceptions2(ValueError, TypeError) 

And notice also how I've added a guard against the empty-call case (except () is invalid at runtime).

Python partial and placeholders

When Python 3.14 was released I had already read about some of its main features (those that involve a PEP and that have been discussed in the Python discussion forums), like Lazy Annotations and Template Strings. When reading in depth recently the release notes I came across a small feature added to functools.partial (and partialmethod) that I find particularly useful:

functools:
Add the Placeholder sentinel. This may be used with the partial() or partialmethod() functions to reserve a place for positional arguments in the returned partial object. (Contributed by Dominykas Grigonis in gh-119127.)

Just a reminder of what partial function application is (don't confuse it with the related concept of curried functions):

In computer science, partial application (or partial function application) refers to the process of fixing a number of arguments of a function, producing another function of smaller arity.

Indeed I already talked about functools.partial some time ago

The "basic" approach to partial function application is that we can just fix (pre-fill) arguments from left to right. This is what we have also in JavaScript with function.prototype.bind (that binds as first argument the "this" value). As Python supports named arguments, functools.partial already supported fixing named arguments.

def format_geo_info(country, region, city, population):
    return f"{city}, {region.upper()} ({country}) - {population}"
    
bound_format = functools.partial(format_geo_info, "France")
print(bound_format("Occitanie", "Toulouse", 500_000))
# Toulouse, OCCITANIE (France) - 500000
print(bound_format("Occitanie", city="Toulouse", population=500_000))
# Toulouse, OCCITANIE (France) - 500000

What was not possible until this version was fixing some intermediate non-named argument, but this is possible since version 3.14 thanks to the Placehodler sentinel value:

format_french_city_with_unknown_population = partial(format_geo_info, "France", Placeholder, Placeholder, 0)
print(format_french_city_with_unknown_population("Ile de France", "Saint Denis"))
# Saint Denis, ILE DE FRANCE (France) - 0

Not a revolutionary feature, but one that I've missed occasionally. A trivial implementation could be something like this:

# supports positional and keyword arguments, but not placeholders
def my_basic_partial(func, *args, **kwargs):
    return lambda *fargs, **fkwargs: func(*args, *fargs, **(kwargs | fkwargs))
    
# add support for placeholders in the arguments
PLACEHOLDER = object()
def my_complete_partial(func, *args, **kwargs):
    def new_func(*fargs, **fkwargs):
        merged_args = []
        fargs_iter = iter(fargs)
        for arg in args:
            if arg is PLACEHOLDER:
                merged_args.append(next(fargs_iter))
            else:
                merged_args.append(arg)
        merged_args.extend(fargs_iter)
        return func(*merged_args, **(kwargs | fkwargs))
    return new_func

format_french_city_with_unknown_population = my_complete_partial(format_geo_info, "France", PLACEHOLDER, PLACEHOLDER, 0)
print(format_french_city_with_unknown_population("Ile de France", "Saint Denis"))
# Saint Denis, ILE DE FRANCE (France) - 0

format_2 = my_complete_partial(format_geo_info, "France", city="Toulouse")
print(format_2("Occitanie", population=500_000))
# Toulouse, OCCITANIE (France) - 500000

In my aforementioned previous post about partial in Python I gave some reasons for using partial over directly trapping the variables with a closure (of course internally partial has to use either closures or a callable class). I've just realised that I was missing the main reason, partial is more semantic.

- Intent-Revealing Code: partial(func, arg) explicitly states your intent to partially apply arguments, improving readability and self-documentation. - Declarative Style: It focuses on the result (a new specialized function) rather than the imperative mechanics of capturing lexical scope.

Lodash, the excellent JavaScript library, also features placeholders in its implemention of partial.

Glibc

Compiling to native code, and furthermore for a Linux system... wow, sounds scary, and very, very far away from what I've been doing in the last decade(s). Well, the thing is that my employer decided some months ago that we had to compile to native code some of our Python applications. It's not something performance related, it's for preventing access to the source code of these applications. We were looking into cython, but we settled on Nuitka, an amazing piece of software that has been serving us so well.

Normally almost every native application compiled for a Linux system has been dynamically linked against glibc. OK, and, what's glibc?

The GNU C Library, commonly known as glibc, is the GNU Project implementation of the C standard library. It provides a wrapper around the system calls of the Linux kernel and other kernels for application use. Despite its name, it now also directly supports C++ (and, indirectly, other programming languages).

So when a Linux native application (using glibc) starts, the dynamic linker (libdl.so) will dynamically load the shared objects (SO, .so files, the equivalent to windows DLL's) needed by the application (like glibc.so) and link the callsites to the functions imported from those libraries.

Obviously glibc evolves over time, so, what about versions? First, what glibc version is installed on my system? You can check the SO's loaded by a running process by doing: lsof -p PID | grep .so. Normally you'll see that it's using: libc.so.6 (in Ubuntu it located here: /usr/lib/x86_64-linux-gnu/libc.so.6). That 6 is not the version number (libc.so.6 is the name for the library since 1997!), the version number is something like 2.XX (2.39 in my ubuntu 24.04). You find it by using: ldd --version

So, what happens if I compile my application in a system with one version of glibc and try to run it in a system with a different version? Well, the situation is quite more fine-grained that I thought. Version numbers are not checked at the glibc level, but at the function level. This is so because glibc uses symbolic versioning

The "Symbol Versioning" Approach (Advanced)

This is what glibc uses. It is also used by heavy-hitters like OpenSSL, Qt, and libgcc.

    The Logic: The filename stays the same (e.g., libc.so.6 has been the name since 1997), but individual functions inside the file are tagged with versions.

    The Result: Multiple versions of the same function can coexist in one file. This allows for extreme backward compatibility without breaking the system every time a single function is updated.
    
    The primary goal of symbol versioning is backward compatibility. It allows a single library file to provide multiple versions of the same function so that:

    Old binaries compiled against v2.10 continue to use the v2.10 implementation.

    New binaries compiled against v2.11 use the new v2.11 implementation.

So multiple versions of the same function live inside glibc, and your binary will dynamically link against the one it was compiled for. And, when does the version number of a function change? Normally it only changes if the function interface (the contract, the ABI) changes, but not if its only its internal implementation that changes. So if we compare symbolic versioning to semantic versioning (SemVer, a more familiar versioning schema), we could say that in symbolic versioning a version change corresponds to a Major version in semantic versioning.

You are exactly right: A new Symbol Version is functionally equivalent to a Major Version bump for that specific function. It signals to the linker that the "Contract" for that specific symbol has changed, and old programs should look for the previous contract elsewhere in the same file.

Notice how symbolic versioning is used for functions inside a library, while semantic versioning (when used) is normally used for libraries.

The glibc version (that one obtained with ldd --version) has no importance in terms of loading the library in memory (the dynamic linker will load libc.so.6 regardless of its "internal" version), the important part is the specific version of each function that we try to link.

I guess when you program in C you are aware of the version of each function that you are using, as you have to adapt your code to the ABI of the function if it has changed, but when that happens behind the scenes, that's quite different. In our case, we just write Python code, and the beautiful Nuitka takes care of transforming it to C and then compiling it to native. So it's Nuitka who takes care of writing the C code in accordance to the function versions inside the glibc in the system. So if then you run that binary in a system with an older glibc version it could happen that your binary is "pointing" to a function with a symbolic version (let's say openEncryptedFile@GLIBC_2.12) higher than the one in the older glibc (let's say openEncryptedFile@GLIBC_2.10) present in the current system, and your application will crash. Basically this means that you have to compile your Python application in a system with a glibc version <= that the glibc version in the target system. It feels odd at first, as the starting point is just the same Python code, and if in one system it can just use openEncryptedFile@GLIBC_2.10 why doesn't it compile it always with that 2.10 even if a bigger version (openEncryptedFile@GLIBC_2.12) is present? Well, that's how things work by default, when compiling, code will be linked to the highest version of that function present in the glibc in the compilation machine.

If you wonder if other .so libraries (SO, ELF libraries) also use symbolic versioning, it depends. For smaller, simpler libraries what is usually used is the SONAME approach, the library (.so file) name changes with each version (this is a coarse grained approach).

Symbolic versioning is the technically superior approach, but it is not the universal standard for all ELF libraries. It depends entirely on the library maintainers and their commitment to long-term ABI stability.

In the Linux ecosystem, there are two primary ways to manage library changes:  
1. The "SONAME" Approach (Common)

Most smaller or simpler libraries use the SONAME mechanism. 
You’ve likely seen files like libfoo.so.1 and libfoo.so.2.

    The Logic: If the developers change the interface, they increment the "Major" version number in the filename itself.  

    The Result: Programs linked against libfoo.so.1 will refuse to start 
    if only libfoo.so.2 is present. This is a "heavy-handed" fix because 
    it requires recompiling every program that uses the library even if 
    the specific function they use didn't actually change.

2. The "Symbol Versioning" Approach (Advanced)

This is what glibc uses. It is also used by heavy-hitters like OpenSSL, Qt, and libgcc.

    The Logic: The filename stays the same (e.g., libc.so.6 has been the name since 1997), 
    but individual functions inside the file are tagged with versions.

    The Result: Multiple versions of the same function can coexist in one file. 
    This allows for extreme backward compatibility without breaking the system every time a single function is updated.

To complete this post, I'll add some useful, related commands:

• To check the SO's used by a given program.
  For a binary on disk: ldd /usr/bin/program_name
  For a running process: lsof -p [PID] | grep '\.so'
• To view the symbols used by a program (the specific functions imported from SO's)
  All imported symbols: nm -Du
  Symbols + Versions: objdump -T | grep '*UND*'
  Only glibc symbols: objdump -T | grep 'GLIBC_'
  Library Version Map: readelf -V
  
• To view the symbols/functions exported by glibc in your system: objdump -T /usr/lib/x86_64-linux-gnu/libc.so.6

Some additional findings related to the last command. For example I want to see the versions of pthread_spin_init present in my glibc: objdump -T /usr/lib/x86_64-linux-gnu/libc.so.6 | grep pthread_spin_init

That gives me:

0000000000a4130 g DF .text 000000000000000d GLIBC_2.34 pthread_spin_init 00000000000a4130 g DF .text 000000000000000d (GLIBC_2.2.5) pthread_spin_ini

Which is very interestinng as it shows us that a symbol version is not a sequential counter for that specific function. Instead, it is a timestamp or a marker of the glibc release that defined that specific version of the function's ABI. From a GPT:

How glibc handles ABI changes with symbol versioning

Original version: Suppose foo() was introduced in GLIBC_2.2.5. That version is tagged as foo@GLIBC_2.2.5.

ABI change in glibc 2.32: If glibc developers change the ABI of foo() in version 2.32 (e.g., change its behavior, arguments, or return type in a way that breaks compatibility), they will:

Keep the old implementation as foo@GLIBC_2.2.5.
Add a new implementation as foo@GLIBC_2.32.

At runtime:

A binary linked against glibc 2.2.5 will request foo@GLIBC_2.2.5, and the dynamic linker will resolve it to the old implementation.
A binary linked against glibc 2.32 will request foo@GLIBC_2.32, and get the new implementation.

This mechanism ensures backward compatibility while allowing glibc to evolve.

Python Sentinel Values

Every now and then we need a flag or sentinel value. A unique value that we can distinguish from the normal values that we are processing and that has a particular meaning. a special, unique value used in programming to signal the end of data processing, a loop, or an operation. For example in my recent post about adding null-safety to a pipe function I was using 2 sentinels/flags: NULL_SAFE and COALESCE.

The essential function a sentinel value has to accomplish is to have a unique identity, so that comparing it by identity (Python: is, JavaScript: ===) with any other value/object in our system has to return false. So the most simple approach is just using a new object for each of our sentinels.

NO_INVEST = object()  # Sentinel value

def invest(amount: int | None | object) -> str:
	if amount is NO_INVEST:
		return "No investment"
	else:
		amount = amount or 0
		return f"we've invested {amount}"

print(invest(NO_INVEST))  # Output: No investment
print(NO_INVEST)  # Output: object object at 0x...

That simple approach works fine, but it's missing a few things. Printing the value is messy (we get an "object object at 0x..." representation, it would be nice to get NO_INVEST) and its rather typing unfriendly. Saying that invest can receive an object, apart from int or None lacks any meaning. What kind of object is that?. A sentinel value should (mainly) have these features:

• A unique identity (is comparison)
• A meaningful repr (debuggability)
• Clear typing (especially for static type checkers)

Our basic sentinel only provides the first one (identity). That's why some smart guy came up with a very interesting PEP 661 proposing a new Sentinel class. Unfortunately that PEP is in deferred status. The document also presents different techniques commonly used for Sentinel values, like the simple object() that I've just shown, using an enum or using a class. Using a class is the best approach to me, I'll show several iterations until getting what I think is the best we can get so far.

Approach 1. Classes are objects in Python. We can use a class for each sentinel object, and in order to get a nice representation we can give it a metaclass with a custom __repr__. The missing piece is having some typing friendliness, we're still stuck with the Any signature. Additionally, declaring a class for something that is not intended to work as an object factory, but to be used as an object in itself is rather unnatural.

    class SentinelMeta(type):
        def __repr__(cls):
            return cls.__name__
        
    class Sentinel(metaclass=SentinelMeta):
        pass
        
    class NO_INVEST(Sentinel): pass


    #def invest(value: int | None | Type[NO_INVEST]) -> str: # this signature feels a bit strange, but it works
    #def invest(value: int | None |Literal[NO_INVEST]) -> str: # this one feels better, but only works with mypy, not with pylance
    def invest(amount: int | None | Any) -> str:
        if amount is NO_INVEST:
            return "No investment"
        else:
            amount = amount or 0
            return f"we've invested {amount}"

    print(invest(NO_INVEST))  # Output: Using NO_INVEST sentinel in match statement
    print(NO_INVEST)  # Output: NO_INVEST

Approach 2. We can make the usage quite more natural by hiding the class creation behind a function. That also allows us to skip the Sentinel base class.

    class SentinelMeta(type):
        def __repr__(cls):
            return cls.__name__

    def sentinel(name: str):
        return SentinelMeta(name, (), {})
    
    NO_INVEST = sentinel("NO_INVEST")


    #def invest(value: int | Type[NO_INVEST]) -> str: # pylance doesn't like it, using a variable as type
    #def invest(value: int | Sentinel) -> str: # we don't have a Sentinel class... so forget it
    def invest(amount: int | None | Any) -> str:
        if amount is NO_INVEST:
            return "No investment"
        else:
            amount = amount or 0
            return f"we've invested {amount}"
        
    print(invest(NO_INVEST))  # Output: Using NO_INVEST sentinel
    print(NO_INVEST)  # Output: NO_INVEST

This one feels quite natural to use, but we still have the problem with typing. We can leverage the Generic types/class subscripting that we saw in my previous post for getting something like this (Approach 3)

    class SentinelMeta(type):
        def __repr__(cls):
            return cls.__name__
    
    class Sentinel(metaclass=SentinelMeta):      
        def __class_getitem__(cls, item):
                return cls
    
    def sentinel(name: str):
        return SentinelMeta(name, (Sentinel,), {})
    
    NO_INVEST = sentinel("NO_INVEST")


    #def invest(value: int | Sentinel) -> str:
    def invest(amount: int | None | Sentinel[NO_INVEST]) -> str: # at a typing level it's equivalent to the above, but it's provides extra meaining      
        if amount is NO_INVEST:
            return "No investment"
        else:
            amount = amount or 0
            return f"we've invested {amount}"

    print(invest(NO_INVEST))  # Output: Using NO_INVEST sentinel      
    print(NO_INVEST)  # Output: NO_INVEST
	

Hey, this one feels rather good to me. The Sentinel[NO_INVEST] looks pretty nice in that signature. The type checking is mainly off, cause our sentinel() function is not typed, so it's considered as returning Any, and when we pass Any to a function it disables type checking. This means that for the type system any sentinel that we create with sentinel() is just an Any, so the type checker will allow passing any sentinel to a function that expects a specific sentinel. It's not a problem for me, what I'm mainly interested in is the semantics, the clarity, that this type annotation provides to the function signature.

Given that we are using Sentinel classes as objects, not as object factories, we can make these classes to be more object like, by preventing instantiation. Additionally, Sentinels do not have attributes and are not intended to be expanded with attributes, so we can prevent them from getting attributes added dynamically. All in all we get:

    class SentinelMeta(type):
        def __repr__(cls):
            return cls.__name__
        
        # prevent sentinel classes from being instantiated
        def __call__(cls, *args, **kwargs):
            raise TypeError(f"{cls.__name__} is a sentinel and cannot be instantiated") 
        
        # prevent sentinel classes from being modified
        def __setattr__(cls, name, value):
            raise AttributeError(f"Cannot modify sentinel {cls.__name__}")


    class Sentinel(metaclass=SentinelMeta):      
        def __class_getitem__(cls, item):
                return cls

    
    def sentinel(name: str):
        return SentinelMeta(name, (Sentinel,), {})

    
    NO_INVEST = sentinel("NO_INVEST")

There's something missing in this implementation, support for our sentinels to be pickled (particularly important if we plan to use them in multiprocessing scenarios). That complicates the design and I've deliberately left it aside for the moment. Maybe we'll see it in another post.