TimestampSigner should use a more precise timestamp

[floguy]

Generating a secure cookie using the TimestampSigner within the same second and the same data actually generates precisely the same cookie.

I'm currently developing a cookie-based session backend, and this line keeps failing as a result, because cycling the cookie isn't really cycling the cookie: https://code.djangoproject.com/browser/django/trunk/django/contrib/sessions/tests.py#L153

I've attached a small patch that fixes the problem by increasing the precision.

[floguy]

Increased precision for TimestampSigner

[Ernesto Rico-Schmidt]

Patch looks good, it applies without problem. The tests for signing and sessions pass.

[floguy]

I've just updated the patch. This new patch adds tests, and actually improves the code's testability in general. Instead of monkeypatching time.time, we pass the time function as an argument to the TimestampSigner class.

[Andrew Godwin]

Looks good to me. Only possible improvement could be renaming self.time to self.time_func (to avoid ambiguity), but that's a very minor change and something the committer can decide on, if needed.

[Jannis Leidel]

time_func should it be.

[Andrew Godwin]

Modified patch to use time_func

[Andrew Godwin]

In [16356]:

Fixed #16182: Increase timestamp precision on TimestampSigner. Thanks to Eric Florenzano.

[Paul McMillan]

I see this is already committed, but I'd like to raise an objection anyway. The timestamp signer was deliberately only using second-level precision. Multiplying by 10,000 significantly increases the length of our sig that we're storing in the already very limited 4095 bytes of the cookie space. We don't want to take up another couple of those when a user can be storing data there.

I'll go into more detail in #16199, but signed cookies like this don't need this functionality for rotation. We're not invalidating any data on the backend, and a previously signed cookie is still technically valid until it expires even if we've rotated it on the users end.

[Jannis Leidel]

Indeed, unfortunately I wasn't involved in the discussion during the DjangoCon sprint and I agree with Paul's argumentation. Reverting seems like the only option to me.

[Andrew Godwin]

Given the arguments against, I'm inclined to agree; even with the new code, it's possible to make two cookies the same, so we're not eliminating any timing problems, merely reducing their probability (and possibly hiding some nasty bugs).

[Andrew Godwin]

In [16425]:

Reverting [16376] in preparation for reverting [16356]. See #16182.

[Andrew Godwin]

Changeset [16426] reverts this commit, and means I'll re-close this ticket as wontfix.