Test failure an arm64, ppc64el and s390x

[lucaskanashiro]

Test failure an arm64, ppc64el and s390x

[LTLA]

FWIW I had similar problems with test outputs changing due to numeric precision issues on different builders/architectures. The most recent example involved differences in rounding for single-precision std::exp() on between Clang and GCC. Another memorable case involved a flip in the branching for a <= if condition between 32-bit Windows and 64-bit OS's.

In the end, I just gave up and wrapped the test in a skip_on_os() call. I concluded that the precision differences weren't going to go away, no matter what I did; but I still needed to test something, so I couldn't just delete the entire test suite.

So, an expedient "solution" might just be to skip the tests based on the architecture. Sys.info()[["machine"]] should tell you whether it's x86_64 or arm64; dunno about the others. Yeah, people might complain about not being able to reproduce the analysis results across machines, but sometimes we can't abstract away differences in the underlying hardware.

[jlmelville]

Many thanks @lucaskanashiro, I am attempting to skip that one test for the architectures listed and if all goes well this will prevent failures in your tests after the next release of this package to CRAN

[jlmelville]

Thanks for the report. I don't have access to any of those platforms (I admit to not even knowing what s390x is) but I'll see what I can do. What are the implications for failing to get this fixed in a timely fashion?

[lucaskanashiro]

@jlmelville the tests should be fixed ASAP to allow the inclusion of the latest version in the next Ubuntu release which will be released next month. For Debian, there is more time since it should be released next year.

FWIW I have access to machines running those architectures, if you need any testing I will be happy to help.

[jlmelville]

Thank you for the offer of testing help, that is appreciated.

The problem here is that I would need to really dig into the code to debug what causes the failures. Even if I can fix it, the chances of a new CRAN release in the next month is zero. So to set expectations and avoid any disappointment, there is no way these can be included in the next Ubuntu. I'm not even sure why someone turned this into a debian package and I can't commit to shouldering the burden of maintaining them. But I will do what I can.

[jlmelville]

@lucaskanashiro do you have the ability to run the R console on those test machines for each failing test architecture and report the value of:

• Sys.info()[["sysname"]]
• R.version$arch

[lucaskanashiro]

The requested info is below:

• arm64

> Sys.info()[["sysname"]]
[1] "Linux"
> R.version$arch
[1] "aarch64"

• ppc64el

> Sys.info()[["sysname"]]
[1] "Linux"
> R.version$arch
[1] "powerpc64le"

• s390x

> Sys.info()[["sysname"]]
[1] "Linux"
> R.version$arch
[1] "s390x"