Bool Mangling in Requests: Asking Forgiveness in Python, Pt 1

The if obj Pattern

I recently ran into the following pattern in some code that implemented the very popular and ubiquitous (for good reason) requests library:

import requests

r = requests.post(url="some.url.tld", headers=some_dict)

if r:
    do_some_stuff(r.json())
else:
    freak_out()

Now, this was weird for me to see. My initial interpretation of what this code was doing was checking for the existence of a requests.Response object from a POST request and if said object exists, do something with that response. In other words: let’s make a POST request, and if everything is fine and dandy, it should return a requests.Response object we can process, and if not, it won’t. So we should check for that object, and continue on if it exists.

Methods and Properties of requests.Response

Seems okay in theory, but I think most users of the requests library are aware that there will be no None object returned in any circumstance, so this if check will always pass and we’ll always do_some_stuff() and never freak_out() even if there’s something wrong with the request. Ideally, we’d use the requests.Response.raise_for_status() method with a try: except: block to properly handle a bad request.

I made a comment about this during code review and decided to do some testing on Python 3.9 to check out this behavior a little more, using a purposefully unauthenticated and improperly formed request to an endpoint of The Cat API for testing:

# test.py

import requests

def if_r():
    r = requests.post("https://api.thecatapi.com/v1/images/upload")
    if r:
        print("if r")

def raise_for_stat():
    r = requests.post("https://api.thecatapi.com/v1/images/upload")
    r.raise_for_status()

if __name__ == '__main__':
    if_r()
    raise_for_stat()    

To my (slight) surprise, I did not get "if r" printed to stdout, only an HTTPError raised. I say “slight” because Python if is of course in reality not an existence check but a truthiness check; it just so happens that bool(None) evaluates to False so you can use if as an existence check…sometimes. I was most surprised that a request.Response object had a change in bool evaluation based on request status. This felt weird to me; the idiom in Python is to ask for forgiveness (handle exceptions) rather than check for validity, as part of being explicit rather than implicit, so it seemed weird you could use requests in a non-idiomatic way, especially given the ease of using requests.Response.raise_for_status().1 I ended up looking at the source code for requests.Response and found something interesting:

def __bool__(self):
    """Returns True if :attr:`status_code` is less than 400.

    This attribute checks if the status code of the response is between
    400 and 600 to see if there was a client error or a server error. If
    the status code, is between 200 and 400, this will return True. This
    is **not** a check to see if the response code is ``200 OK``.
    """
    return self.ok

So the __bool__ special method is just a wrapper around the requests.Response.ok property, which shows up explicitly in the API documentation for requests.Response. requests.Response.ok probably just runs some checks on the status_code attr then, right?

@property
def ok(self):
    """Returns True if :attr:`status_code` is less than 400, False if not.

    This attribute checks if the status code of the response is between
    400 and 600 to see if there was a client error or a server error. If
    the status code is between 200 and 400, this will return True. This
    is **not** a check to see if the response code is ``200 OK``.
    """
    try:
        self.raise_for_status()
    except HTTPError:
        return False
    return True

Nope! It just uses the good ol’ try: except: control flow with raise_for_status() Considering this is what I was advocating to use, I was very pleasantly surprised to find this pattern in the source code.

Why This Irritates Me

In my opinion, using if r: like this becomes a very roundabout way of avoiding exception handling without avoiding exception handling at all. To avoid using try: except:, you’d call bool(some_response_object) which looks at some_response_object.ok property…which uses try: except:. One of the reasons to avoid exception handling is the performance cost of throwing and handling an exception, but in this case you’re going to throw and handle an exception no matter what since that’s how self.__bool__/self.ok is implemented here. Plus, exceptions aren’t even that expensive in Python to begin with, so you’re left with readability and complexity claims. Sure, there are definitely cases where readability will suffer due to extended try: except: blocks, especially if these are located within other control flows, but I’d argue if r: isn’t that readable to begin with, since, if you don’t know that self.__bool__ in this case just shadows self.ok, which I personally did not until I looked at the source code, it can be ambiguous on how the truthiness of the requests.Response object is determined, and in which cases the if block will evaluate as True.

I am also just not sure in general how much I like the fact that the truthiness of a requests.Response instance is dependent on an external system. Perhaps it’s my slight affinity for functional programming and formal logic (Haskell was the first language I ever touched, fun fact!) that makes me a little adverse to this; it just feels off to me to rely on either internal or external state for truth value testing in this way, even though epistemically I hold that truthiness can be evaluated in terms of external relations to the world in the case of empirical statements (I will stop here in regards to epistemology since it would take several book-length works for me to satisfactorily talk about it). The Python People™ definitely don’t share quite the same reservations; if object.__bool__(self) is undefined __len__() will be called instead and non-zero values will be considered true, meaning there is a strong relationship between state and truthiness (if __len__() is undefined __bool__() then becomes an existence check). Still, I do think relying on properties or attrs of objects for control flow like this, rather than how truthiness is implemented, is better for readability and explicitness.

Some Final Thoughts

Overall, I find requests.Response.__bool__() to be an example of what I call bool mangling, a not-very-explicit-or-clear-on-first-glance implementation of __bool__ that may be intended as sugar but doesn’t really do much for readability, at least in my opinion. Instead of using if r:, we should ask for forgiveness, use r.raise_for_status(), and properly handle HTTP errors as they occur when we’re making requests, rather than just suppressing them and examining the validity of a response using truth value testing. It’s better to clearly communicate what the code is doing (e.g. processing data only on a successful request) than not; as The Zen of Python states, “Explicit is better than implicit.” And, when we do need to use an if check in this case, in the times when the r.raise_for_status() pattern becomes too explicit (“Readability counts.”), we should use if r.ok: instead, to clearly communicate what we’re doing.


  1. Funnily enough I later realized that in the docs for requests.Response.json() the Requests maintainers explictly say to use raise_for_status() instead of ugly if checks, since if r.json() has slightly different behavior from its parent class. ↩︎