refactor(r): Refactor ArrowArray(Stream) -> R Vector conversion #392
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #392 +/- ##
==========================================
- Coverage 88.74% 86.09% -2.66%
==========================================
Files 81 97 +16
Lines 14398 15091 +693
==========================================
+ Hits 12778 12993 +215
- Misses 1620 2098 +478 ☔ View full report in Codecov by Sentry. |
paleolimbot
force-pushed
the
r-convert-refactor
branch
from
b44feef to
6598171
Compare
paleolimbot
force-pushed
the
r-convert-refactor
branch
from
2afc7b8 to
a749d8d
Compare
paleolimbot
added a commit
that referenced
this pull request
May 15, 2024
…461) This PR adds the `nanoarrow_vctr`, which is an R translation of the Python `Array` class in nanoarrow's Python bindings. This is implemented like an R `factor()` in the sense that under the hood it is a sequence of integers (`0:(array$length - 1)` at the beginning) with attributes that give those integers context. This is implemented in such a way that it is "tacked on" to the existing conversions. The existing conversions do need a refactoring ( #392 ), but that is a heavy change for this point in the release cycle. The only change needed to the existing conversion was a slight refactor of the "consume array stream" code that correctly gave each array in the stream its own R object to manage its lifecycle (before each array was "materialized" and then immediately released because no previous conversion code required an ArrowArray to live beyond the conversion. The motivation for this change is converting GeoArrow extension types. In the geoarrow package, we implement an efficient conversion from a stream of arrays to various types of R-spatial objects (e.g., sf); however, we really don't want to invoke the default conversion for those types because they have awful performance (e.g., the multipolygon would be a `list(list(list(data.frame))))`) and there's no need to invoke that number of R object conversions between the initial state (an arrow array) and the final state (an sfc column). The nanoarrow_vctr allows something like: ```r df <- convert_array(some_array_containing_a_geoarrow_col) st_as_sfc(df$geometry) # or s2::as_s2_geography(df$geometry), or something else ``` A side-effect of this change is that we have an escape hatch for conversions that are lossy or contain types with no R equivalent. A quick demo: ``` r library(nanoarrow) arrays <- lapply( list(1:5, 6:10, 11:13), as_nanoarrow_array ) # A vctr can be created from any stream (vctr <- as_nanoarrow_vctr(basic_array_stream(arrays))) #> <nanoarrow_vctr int32[13]> #> [1] 1 2 3 4 5 6 7 8 9 10 11 12 13 # Under the hood this is something like a factor() where levels are # a list of arrays with cached offsets. This is like an Arrow ChunkedArray str(vctr) #> <nanoarrow_vctr int32[13]> #> List of 3 #> $ :<nanoarrow_array int32[5]> #> ..$ length : int 5 #> ..$ null_count: int 0 #> ..$ offset : int 0 #> ..$ buffers :List of 2 #> .. ..$ :<nanoarrow_buffer validity<bool>[0][0 b]> `` #> .. ..$ :<nanoarrow_buffer data<int32>[5][20 b]> `1 2 3 4 5` #> ..$ dictionary: NULL #> ..$ children : list() #> $ :<nanoarrow_array int32[5]> #> ..$ length : int 5 #> ..$ null_count: int 0 #> ..$ offset : int 0 #> ..$ buffers :List of 2 #> .. ..$ :<nanoarrow_buffer validity<bool>[0][0 b]> `` #> .. ..$ :<nanoarrow_buffer data<int32>[5][20 b]> `6 7 8 9 10` #> ..$ dictionary: NULL #> ..$ children : list() #> $ :<nanoarrow_array int32[3]> #> ..$ length : int 3 #> ..$ null_count: int 0 #> ..$ offset : int 0 #> ..$ buffers :List of 2 #> .. ..$ :<nanoarrow_buffer validity<bool>[0][0 b]> `` #> .. ..$ :<nanoarrow_buffer data<int32>[3][12 b]> `11 12 13` #> ..$ dictionary: NULL #> ..$ children : list() # vctrs can be sliced: head(vctr) #> <nanoarrow_vctr int32[6]> #> [1] 1 2 3 4 5 6 # ...and can live in a data.frame head(tibble::tibble(x = vctr)) #> # A tibble: 6 × 1 #> x #> <nnrrw_vc> #> 1 1 #> 2 2 #> 3 3 #> 4 4 #> 5 5 #> 6 6 # They can be used as zero-copy conversion targets array <- as_nanoarrow_array(1:5) convert_array(array, nanoarrow_vctr()) #> <nanoarrow_vctr int32[5]> #> [1] 1 2 3 4 5 # ...also works in a nested ptype array <- as_nanoarrow_array(data.frame(x = 1:5)) convert_array(array, tibble::tibble(x = nanoarrow_vctr())) #> # A tibble: 5 × 1 #> x #> <nnrrw_vc> #> 1 1 #> 2 2 #> 3 3 #> 4 4 #> 5 5 # For nested list output, it will give a slice of the original array for # each list item array <- as_nanoarrow_array( list(1:5, 6:10, NULL, 11:13), schema = na_list(na_int32()) ) (lst_of <- convert_array(array, vctrs::list_of(nanoarrow_vctr()))) #> <list_of<nanoarrow_vctr>[4]> #> [[1]] #> <nanoarrow_vctr int32[5]> #> [1] 1 2 3 4 5 #> #> [[2]] #> <nanoarrow_vctr int32[5]> #> [1] 6 7 8 9 10 #> #> [[3]] #> NULL #> #> [[4]] #> <nanoarrow_vctr int32[3]> #> [1] 11 12 13 for (i in seq_along(lst_of)) { array <- attr(lst_of[[i]], "chunks")[[1]] cat(sprintf("offset: %d, length: %d\n", array$offset, array$length)) } #> offset: 0, length: 5 #> offset: 5, length: 5 #> offset: 10, length: 3 ``` <sup>Created on 2024-05-10 with [reprex v2.1.0](https://reprex.tidyverse.org)</sup>
paleolimbot
force-pushed
the
r-convert-refactor
branch
from
a749d8d to
15b38b6
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
The initial version of converting ArrowArray (or stream of them) to R is implemented in C and is difficult to understand. Not only is this difficult because of the verbose C, the dispatch portion is implemented almost completely twice (once for a single array, once for an array stream). It is at a point currently where it is difficult for me, let alone an external contributor, to add features or fix bugs. Time to refactor!
This approach uses C++ classes/virtual method dispatch to handle the different types of vector conversions. This is similar to how the arrow R package does this except the Arrow R package uses the Arrow C++ converter infrastructure/heavy templating to do dispatch. Here we use a
switch()and eat the per-batch and per-column virtual method call.Work in progress!