Non-precise test cases / comparing numbers

google / open-location-code

Open Location Code is a library to generate short codes, called "plus codes", that can be used as digital addresses where street addresses don't exist.

https://plus.codes

Apache License 2.0

4.06k stars 471 forks source link

Non-precise test cases / comparing numbers #448

Closed fulldecent closed 2 months ago

fulldecent commented 3 years ago

When creating test cases, it is best practice to NOT compare floating point values using an equality check.

Therefore, test cases like these are inappropriate: https://github.com/google/open-location-code/blob/master/test_data/decoding.csv#L26

Or, there should be a lot more documentation (in the test case) to explain why that specific value is expected.

References:

Best practice: https://floating-point-gui.de/errors/comparison/
For just CPP, Bazel, the preferred testing framework here for CPP offers the following https://google.github.io/googletest/advanced.html#floating-point-macros

drinckes commented 2 months ago

I'm not super happy with having floating point checks but we have to try to define a set of test cases that function across different languages with different precision capabilities and this is the best we came up with.

If you want to propose an alternative feel free. Otherwise I don't see this as something we need to address.

fulldecent commented 2 months ago

I did have the macros there as a proposed solution.

When doing a decoding test, we should not require the result to exactly equal the required test case.

Instead we should only require them to be within 0.00000001 of each other. Because on 32-bit systems numbers are considered equivalent if they are that close.

This issue can please stay open. I can contribute PRs. Working on this is blocked on updating the specification. Specifically if we are still supporting 32-bit implementations or only 64-bit. (Those will have slightly different ways to properly do the test case.)