Custom conversion function no longer works with reticulate >= 1.36.0

[lazappi]

Hi

In the {zellkonverter} Bioconductor package we have some custom Python to R conversion functions. One of these that is designed to handle numpy recarray's (array type 20) no longer works with {reticulate} >= 1.36.0.

Here is a small example (there is probably a nicer way to write this but this works):

library(zellkonverter)

reticulate::py_run_string("import numpy as np", convert = FALSE)
reticulate::py_run_string(
    "array = np.array([(1.0, 2), (3.0, 4)], dtype=[('x', float), ('y', int)])",
    convert = FALSE
)
result <- reticulate::py_run_string("rec_array = array.view(np.recarray)", convert = FALSE)

reticulate::py_to_r(result$rec_array)
# With reticulate <= 1.35.0 this works
#> [,1] [,2]
#> [1,]    1    2
#> [2,]    3    4

# With reticulate >= 1.36.0 there is an error
#> Error: Conversion from numpy array type 20 is not supported

As far as I can tell, the custom conversion function is not being called at all. It seems like there were some fairly major changes with v1.36.0 so if you have any advice on what might be causing this change and how to address it that would be great. Or if there is an alternative/better way to handle converting these arrays that would also be helpful.

Thanks!

[t-kalinowski]

Thanks for reporting! This is fixed on main now.

By the way, I noticed that numpy recarrays inherit from S3 class numpy.recarray. It may make sense to narrow the registered S3 method from numpy.ndarray to numpy.recarray.

[lazappi]

Thanks for reporting! This is fixed on main now.

Thanks! Do know when it might reach CRAN?

By the way, I noticed that numpy recarrays inherit from S3 class numpy.recarray. It may make sense to narrow the registered S3 method from numpy.ndarray to numpy.recarray.

Ok. I think that wasn't the case when I wrote this originally but I'll check again. Thanks for the tip.

[t-kalinowski]

We just submitted a new release to CRAN last week (May 21). CRAN policy states that updates should be “no more than every 1–2 months”. Adhering to the CRAN policy, the soonest next update would be ~ June 21. However, occasionally, we do a rapid release in violation of the CRAN policy if circumstances warrant it.

I think this regression has been live on CRAN for ~50 days already (since Reticulate 1.36). How impactful would another 20 days be?

[lazappi]

I think that should be fine in this case as there is a fairly easy workaround by getting people to install v1.35.0/devel if needed. Thanks!