Show examples of building Arrow C++ structures

[wjones127]

For users coming from Arrow C++ / PyArrow, it might not be obvious what nanoarrow structures to create so they can export. Depending on the context, an ArrowSchema can represent a data type, a field, or a schema. Similarly, an ArrowArray can represent an array or record batch (tabular). To start, we should show the correspondence between Nanoarrow structs and Arrow C++ types.

Nanoarrow	Arrow C++
ArrowArray	arrow::Array
ArrowArray where type is struct	arrow::RecordBatch
std::vector<ArrowArray>	arrow::ChunkedArray
std::vector<ArrowArray> where type is struct	arrow::Table
ArrowSchema	arrow::DataType
ArrowSchema	arrow::Field
ArrowSchema	arrow::Schema

Then we may also want recipes for:

• How to build a struct array in nanoarrow, and export as an arrow::RecordBatch
• How to build a record batch reader, and export as an arrow::Table.

[paleolimbot]

That's a great idea! That mapping is definitely not obvious unless you've spent a lot of time with the C data interface.

[eddelbuettel]

I was wondering if you we could bring this back to the fore? I am currently loosing my mind as a simple 'call from R' wrapper for the (nice) linesplitter example works just fine when I start in R and pass an external pointer down:

// Plain Interface
// [[Rcpp::export]]
bool linesplit_from_R_plain(const std::string str, SEXP sxparr) {
    // We get an R-created 'nanoarrow_array', an S3 class around an external pointer
    if (!Rf_inherits(sxparr, "nanoarrow_array"))
        Rcpp::stop("Expected class 'nanoarrow_array' not found");

    // It is a straight up external pointer so we can use R_ExternalPtrAddr()
    struct ArrowArray* arr = (struct ArrowArray*)R_ExternalPtrAddr(sxparr);

    auto res = linesplitter_read(str, arr);
    return true;
}

But when I want to start from C++, I get lost somewhere build the nanoarrow-compliant external pointer up by itself. It all works fine when I cheat (as the nanoarrow package also does in places) and call into R:

// res <- linesplit_from_cpp("the\nquick\nbrown\nfox");
// print(res);
// print(arrow::Array$create(res))
//
// [[Rcpp::export]]
Rcpp::XPtr<ArrowArray> linesplit_from_cpp(const std::string str) {
    Rcpp::Environment ns = Rcpp::Environment::namespace_env("nanoarrow");
    Rcpp::Function f1 = Rcpp::Function("nanoarrow_array_init", ns);
    Rcpp::Function f2 = Rcpp::Function("na_string", ns);
    auto sxparr = f1(f2());

    // It is an external pointer we can access, here with checking
    auto arr = xptr_get_ptr<ArrowArray>(sxparr, "nanoarrow_array");
    auto res = linesplitter_read(str, arr);

    auto s = Rcpp::XPtr<ArrowArray>{sxparr};
    // setting a tag somehow upsets the Arrow nature of things
    //xptr_set_tag(s, Rcpp::wrap(XPtrTagType<ArrowArray>));
    return s;
}

(Apologies for the small bits of Rcpp but that's I am most familiar with. The helpers used above are minimal wrappers around the C API of R as eg the following.)

template <typename T> Rcpp::XPtr<T> make_xptr(SEXP p) {
    return Rcpp::XPtr<T>(p); 	// the default Rcpp::XPtr ctor with deleter on and tag and prot nil
}

So far so good but I still have two problems. I can't seem to build an external pointer 'up from the C/C++ bases' to make nanoarrow happy. It comes out ok, but invoking any of the R-level helpers goes astray. A simple print(res) or print(str(res)) of the return object gets an error of a failing allocation of a bazillion bytes. The other (smaller) problem is that I good some use out of setting (and checking) external pointer tags, somehow doing that here with the one gotten by calling into R also dirties some other bit.

By now there are nice bits and pieces of nanoarrow use now in the adbc* packages (and the repo), in duckdb-r, and possibly also in the geospatial apps (haven't looked). The (vendorable) nanoarrow.* are fine just C/C++ work. The R interface is fine for work at the REPL prompt. But there is much less on 'how to work with nanoarrow for R extensions' (and ditto for Python where support is slowly growing in the package). So would there be some appetite to extend, say, what is in nanoarrow.hpp in light of possible 'interface helpers' ? I'd be happy to help along, of course.

PS And the reason I put this here into the 'more C++ examples' issue was that I do enjoy the bits and pieces of debugging infra. I have some alternate uses of (nano)arrow but am thinking of rewriting parts of those to stay closer to nanoarrow to take advantage of some that debug infra. And when I try to remain 's3-classed' weird things happen that do not happen if export just plainer external pointers closer to the plain void* pointer pair.

[paleolimbot]

setting a tag somehow upsets the Arrow nature of things

Yes: the tag is where (perhaps inadvisably) the external pointer to the ArrowSchema is stored, and I'm not sure the class is checked everywhere so perhaps the failing allocation of a bazillion bytes is because it's being misinterpted somehow. The protected member is where you should set any SEXP dependency, with the caveat that if there is already one there you have to maintain a reference to it (e.g., list(your_new_sexp_dep, old_sexp_dep)). You can use nanoarrow_pointer_export() to wrap the array in another array that maintains the reference via the release() callback instead of via the prot tag. All of that is undocumented, of course...I didn't expect this level of internal use quite yet but obviously it should be clear 🙂 .

So would there be some appetite to extend, say, what is in nanoarrow.hpp in light of possible 'interface helpers' ?

For R-specific helpers, perhaps a header like the one you mentioned in r/inst/include/nanoarrow/nanoarrow_r.h|hpp? Even if all they do is Rf_eval() to call into R in the first iteration (we can make them faster later if the call into R is limiting). Usually I just allocate from R and pass the array_xptr into the C/C++ function (e.g., https://github.com/apache/arrow-adbc/blob/main/r/adbcdrivermanager/R/adbc.R#L180-L191 ). At the very least, a copy of the C Data/Stream structures would be helpful.

For Python, we generate Cython definitions (nanoarrow_c.pxd), and I've wondered if it's worth putting that in dist/ (Python extensions can just copy nanoarrow.h, nanoarrow.c, and nanoarrow_c.pxd and it's reasonably easy to wrap from there (Cython is considerably easier as a glue language to anything we have in R). This is what I've done in https://github.com/geoarrow/geoarrow-c/tree/main/python . nanoarrow for Python will probably serve a similar purpose (help with the allocating). In Python there is also https://arrow.apache.org/docs/format/CDataInterface/PyCapsuleInterface.html , which is more or less what as_nanoarrow_XXX() is trying to do.

For C++ (i.e., nanoarrow.hpp), there is almost certainly more that could be useful, although I am hesitant to increase the scope beyond not leaking memory. It might be worth drafting an internal set of helpers you developed and used somewhere and linking to it here? (Or maybe that's not what you had in mind)

'how to work with nanoarrow for R extensions'

This should definitely be a vignette/article! As you noted there is now ADBC and soon geoarrow (and whatever you are up to!) which are the first few test cases. Porting the linesplitter example would be a great place to start, as you noted (i.e., here's how you'd wrap the function in an R package...).

[eddelbuettel]

Thanks for the quick reply! I am definitely up for brainstorming a little more and 'slowly but surely' expanding this. I'll also try to clean up my trivial little wrapper around linesplitter so that we have a base to start. Currently fighting other fires but interesting to see you say "Usually I just allocate from R". I had reasonably good luck allocating directly (as per earlier exchanges and tickets) and have gotten that mostly by valgrind and other checkers. (It's a bit bizarre: all test script pass in isolation ie running one at a time, but when I check the package someone somewhere is unhappy and not cleaned. As you say in a dry comment "almost impossible" to check without some framework so I am happy to lean on what you have here ... but I am stumbling a little at the gate.)

(And while I don't do much Python colleagues do and having "moar Python" with nanoarrow would make it easier for us in general. So thumbs up there too.)

[eddelbuettel]

Yes: the tag is where (perhaps inadvisably) the external pointer to the ArrowSchema is stored,

Dang. Re-reading this sentence I think I once knew that too from reading your code but forgot. That explains that part. I can try to also put one there when I build up 'from the other direction'.

[eddelbuettel]

Tried setting the schema external pointer as a tag, but no luck so far.

I put a (truly minimal) example up here. It builds and checks cleanly for me, and has one example with either nanoarrow (required) or arrow (suggested) export:

> example(linesplit, package="linesplitter")

lnsplt> txt <- "the\nquick\nbrown\nfox"

lnsplt> linesplit(txt)
<nanoarrow_array string[4]>
 $ length    : int 4
 $ null_count: int 0
 $ offset    : int 0
 $ buffers   :List of 3
  ..$ :<nanoarrow_buffer validity<bool>[0][0 b]> ``
  ..$ :<nanoarrow_buffer data_offset<int32>[5][20 b]> `0 3 8 13 16`
  ..$ :<nanoarrow_buffer data<string>[16 b]> `thequickbrownfox`
 $ dictionary: NULL
 $ children  : list()

lnsplt> linesplit(txt, TRUE)
Array
<string>
[
  "the",
  "quick",
  "brown",
  "fox"
]
> 

[paleolimbot]

Cool! Is the example somewhere I can find it!? I would love to add it to examples/linesplitter or in exploded from to r/vignettes/articles/extending.Rmd (or blog post, or whatever!)

Tried setting the schema external pointer as a tag, but no luck so far.

Hmm...I also usually do this in R (nanoarrow::nanoarrow_array_set_schema()). It seems pretty straightforward but I definitely may have missed something! https://github.com/apache/arrow-nanoarrow/blob/main/r/src/array.h#L80-L88

[eddelbuettel]

The example I show above is in the manual page (and in the R code via the usual means). There are a few things I should clean up still -- I like have types other than SEXP in the signature, that requires the usual inst/include/PACKAGENAME_types.h but I realized I probably don't want all of nanoarrow.h there but maybe just the standard C API of Arrow part.

(As an aside you have a very very large part of nanoarrow hidden behind static which also leads the duckdb use to copy functions out. That ... can't really be ideal so maybe we can talk about 'liberating' some of those functions.)

Otherwise happy to stick the example into the README / add a quick tinytest test predicate or two.

Ah, and in all the excitement I promptly forgot to point to the repo I made (quickly) for this: https://github.com/eddelbuettel/linesplitter

[paleolimbot]

As an aside you have a very very large part of nanoarrow hidden behind static which also leads the duckdb use to copy functions out

I didn't know duckdb was using it! All static functions are internal and could change at any time (although I'm happy to expose functionality if there's something missing that can't be accessed by the functions in nanoarrow.h).

[eddelbuettel]

Yeah it's in https://github.com/duckdb/duckdb-r/tree/main/src/duckdb/src/common/adbc