Thanks for putting this together -- I've been on-and-off following under the different names it had over the last few months.

One thing I noticed, and which may be due to my use off gcc / g++ (now at -10 or -11) with you presumably using clang+= is that I get two nuisance warnings. They are easy to fix and I would be happy to send you a PR:

• in src/cast-from-array.c you declare double_copy_result to hold the function return, but the compiler is unhappy that the result is never used -- simple fix is to not declare and assing
• in src/narrow/buffer.cc we have simple signed/unsigned mismatch.

That's all, and I don't mean to intrude or trample of your feet but you appear to be a rather careful coder so maybe you'd value two fewer warnings.

Your have

# only needed for arrow package before 7.0.0
<- function(ptr) {
  .Call(narrow_c_pointer_addr_dbl, ptr)
}

and I had been / am using schema and array pointers cast to double which maps reasonably well for non-R use. Is there a reason this could not be exported, whether or not Arrow 7.0 changed this or not. (Also, changed to ... what? I did not immediately find a replacement or alternate in the updated arrow package but I will admit that I do find navigating its code base a little difficult given the size and rate of change).

Also, #5 really asked for a C-level API, as helpful as the exported R functions are. I may still play with a local variant to export the object code and see how it goes.

Thanks again for the package. It is quite helpful.

...probably a misunderstanding on my part about the object types that are expected:

library(carrow)

  # some test data
  df <- data.frame(a = 1L, b = 2, c = "three")
  batch <- arrow::record_batch(df)
  tf <- tempfile()

  # write a valid file
  file_obj <- arrow::FileOutputStream$create(tf)
  writer <- arrow::RecordBatchFileWriter$create(file_obj, batch$schema)
  writer$write(batch)
  writer$close()
  file_obj$close()

  # create the reader
  read_file_obj <- arrow::ReadableFile$create(tf)
  reader <- arrow::RecordBatchFileReader$create(read_file_obj)

  # export it to carrow
  stream <- as_carrow_array_stream(reader)

  schema <- carrow_array_stream_get_schema(stream)
  identical(
    carrow_schema_info(schema, recursive = TRUE),
    carrow_schema_info(as_carrow_schema(reader$schema), recursive = TRUE)
  )
#> [1] TRUE

  # skip("Attempt to read batch from exported RecordBatchReader segfaults")
  # batch <- carrow_array_stream_get_next(stream)

  read_file_obj$close()
  unlink(tf)

^{Created on 2021-11-23 by the reprex package (v2.0.1)}

narrow-check.zip shows what happens when CMD check on Fedora 36 gcc 12.1.1 and the Fedora libarrow etc. binaries are used. Same for geoarrow: geoarrow-check.zip

The GDAL 3.5.0 Arrow and Parquet drivers read the GDAL autotest/ogr/data examples OK - my reason for raising a diffuse issue is https://github.com/wcjochem/sfarrow/issues/14, a new failing revdep check on F36/gcc12 and Fedora's arrow RPMs (no external arrow libraries at all previously). I think, since you are obviously looking at checking the GDAL drivers, it might be worth seeing how to cross-check sf with GDAL 3.5.0 and the two experimental drivers, geoarrow and sfarrow (and others if they appear - maybe Python and reticulate). What do you think (@edzer for reference, this format feels promising)?

narrow already automates creating the needed instantiations via src/init.c. That registers the object code.

How do you think about going one step further and provide a file, say, narrowAPI.h with a bunch of declarations such as (untested, adapting from prior work)

SEXP attribute_hidden narrow_c_array_info(SEXP x) {
  static SEXP(*fun)(SEXP) = (SEXP(*)(SEXP)) R_GetCCallable("narrow","narrow_c_array_info");
  return fun(x);
}

and so on.

A number of packages do this (e.g. nloptr, I have two small packages exporting some of R's own (unexported) API functions (via copies in the package), also added it in small scope other packages). The canonical example (also in WRE) is always lme4 and Matrix.

Would you have any interest in supporting such use?