kevcaz / duckdbfs

Simple wrappers for duckdb to open local and remote filesystems

duckdbfs

duckdbfs is a simple wrapper around the duckdb package to facilitate working with the construction of a single lazy table (SQL connection) from a set of file paths, URLs, or S3 URIs.

Installation

You can install the development version of duckdbfs from GitHub with:

# install.packages("devtools")
devtools::install_github("cboettig/duckdbfs")

Quickstart

library(duckdbfs)
library(dplyr)
#> 
#> Attaching package: 'dplyr'
#> The following objects are masked from 'package:stats':
#> 
#>     filter, lag
#> The following objects are masked from 'package:base':
#> 
#>     intersect, setdiff, setequal, union

Imagine we have a collection of URLs to files we want to combine into a single tibble in R. The files could be parquet or csv, and some files may have additional columns not present in other files. The combined data may be very large, potentially bigger than available RAM or slow to download completely, but we may only want a subset using methods like dplyr::filter() or dplyr::summarise().

base <- paste0("https://github.com/duckdb/duckdb/raw/main/",
               "data/parquet-testing/hive-partitioning/union_by_name/")
f1 <- paste0(base, "x=1/f1.parquet")
f2 <- paste0(base, "x=1/f2.parquet")
f3 <- paste0(base, "x=2/f2.parquet")
urls <- c(f1,f2,f3)

We can easily access this data without downloading by passing a vector of URLs. Note that if schemas (column names) do not match, we must explicitly request duckdb join the two schemas. Leave this as default, FALSE when not required to achieve much better performance.

ds <- open_dataset(urls, unify_schemas = TRUE)
ds
#> # Source:   table<csnyzuoxobayjyy> [3 x 4]
#> # Database: DuckDB 0.8.1 [unknown@Linux 5.17.15-76051715-generic:R 4.3.1/:memory:]
#>       i     j x         k
#>   <int> <int> <chr> <int>
#> 1    42    84 1        NA
#> 2    42    84 1        NA
#> 3    NA   128 2        33

Use filter(), select(), etc from dplyr to subset and process data – any method supported by dbpylr. Then use dplyr::collect() to trigger evaluation and ingest results of the query into R.

S3-based access

We can also access remote data over the S3 protocol. An advantage of S3 is that unlike https, it can discover all files in a given folder, so we don’t have to list them individually. This is particularly convenient for accessing large, partitioned datasets, like GBIF: (nearly 200 GB of data split across more than 2000 parquet files)

parquet <- "s3://gbif-open-data-us-east-1/occurrence/2023-06-01/occurrence.parquet"
duckdb_s3_config()
gbif <- open_dataset(parquet, anonymous = TRUE, s3_region="us-east-1")

The additional configuration arguments are passed to the helper function duckdb_s3_config() to set access credentials and configure other settings, like alternative endpoints (for use with S3-compliant systems like minio). Of course it also possible to set these ahead of time by calling duckdb_s3_config() directly. Many of these settings can also be passed along more compactly using the URI query notation found in the arrow package. For instance, we can request anonymous access to a bucket on an alternative endpoint as:

efi <- open_dataset("s3://anonymous@neon4cast-scores/parquet/aquatics?endpoint_override=data.ecoforecast.org")

Spatial data

duckdb can also understand a wide array of spatial data queries for spatial vector data, similar to operations found in the popular sf package. Most spatial query operations require an geometry column that expresses the simple feature geometry in duckdb’s internal geometry format (nearly but not exactly WKB). A common pattern will first generate the geometry column from raw columns, such as latitude and lognitude columns, using the duckdb implementation of the a method familiar to postgis, ST_Point:

spatial_ex <- paste0("https://raw.githubusercontent.com/cboettig/duckdbfs/",
                     "main/inst/extdata/spatial-test.csv") |>
  open_dataset(format = "csv") 

spatial_ex |>
  mutate(geometry = ST_Point(longitude, latitude)) |>
  to_sf()
#> Simple feature collection with 10 features and 3 fields
#> Geometry type: POINT
#> Dimension:     XY
#> Bounding box:  xmin: 1 ymin: 1 xmax: 10 ymax: 10
#> CRS:           NA
#>    site latitude longitude      geometry
#> 1     a        1         1   POINT (1 1)
#> 2     b        2         2   POINT (2 2)
#> 3     c        3         3   POINT (3 3)
#> 4     d        4         4   POINT (4 4)
#> 5     e        5         5   POINT (5 5)
#> 6     f        6         6   POINT (6 6)
#> 7     g        7         7   POINT (7 7)
#> 8     h        8         8   POINT (8 8)
#> 9     i        9         9   POINT (9 9)
#> 10    j       10        10 POINT (10 10)

Recall that when used against any sort of external database like duckdb, most dplyr functions like dplyr::mutate() are being transcribed into SQL by dbplyr, and not actually ever run in R. This allows us to seamlessly pass along spatial functions like ST_Point, despite this not being an available R function. The to_sf() coercion will parse its input into a SQL query that gets passed to duckdb, and the return object will be collected through sf::st_read, returning an (in-memory) sf object.

Note that we can add arbitrary spatial functions that operate on this geometry, provided we do so prior to our call to to_sf. For instance, here we first create our geometry column from lat/lon columns, and then compute the distance from each element to a spatial point:

spatial_ex |> 
  mutate(geometry = ST_Point(longitude, latitude)) |>
  mutate(dist = ST_Distance(geometry, ST_Point(0,0))) |> 
  to_sf()
#> Simple feature collection with 10 features and 4 fields
#> Geometry type: POINT
#> Dimension:     XY
#> Bounding box:  xmin: 1 ymin: 1 xmax: 10 ymax: 10
#> CRS:           NA
#>    site latitude longitude      dist      geometry
#> 1     a        1         1  1.414214   POINT (1 1)
#> 2     b        2         2  2.828427   POINT (2 2)
#> 3     c        3         3  4.242641   POINT (3 3)
#> 4     d        4         4  5.656854   POINT (4 4)
#> 5     e        5         5  7.071068   POINT (5 5)
#> 6     f        6         6  8.485281   POINT (6 6)
#> 7     g        7         7  9.899495   POINT (7 7)
#> 8     h        8         8 11.313708   POINT (8 8)
#> 9     i        9         9 12.727922   POINT (9 9)
#> 10    j       10        10 14.142136 POINT (10 10)

For more details including a complete list of the dozens of spatial operations currently supported and notes on performance and current limitations, see the duckdb spatial docs

Local files

Of course, open_dataset() can also be used with local files. Remember that parquet format is not required, we can read csv files (including multiple and hive-partitioned csv files).

write.csv(mtcars, "mtcars.csv", row.names=FALSE)
lazy_cars <- open_dataset("mtcars.csv", format = "csv")

Limitations

• NOTE: at this time, the duckdb httpfs file system extension in R does not support Windows.

Mechanism / motivation

This package simply creates a duckdb connection, ensures the httpfs and spatial extensions are installed if necessary, sets the S3 configuration, and then constructs a VIEW using duckdb’s parquet_scan() or read_csv_auto() methods and associated options. It then returns a dplyr::tbl() for the resulting view. Though straightforward, this process is substantially more verbose than the analogous single function call provided by arrow::open_dataset() due mostly to the necessary string manipulation to construct the VIEW as a SQL statement. I’ve used this pattern a lot, especially when arrow is not an option (http data) or has substantially worse performance (many S3 URIs).

Advanced notes

This is very similar to the behavior of arrow::open_dataset(), with a few exceptions:

• at this time, arrow does not support access over HTTP – remote sources must be in an S3 or GC-based object store.
• With local file system or S3 paths, duckdb can support “globbing” at any point in the path, e.g. open_dataset(data/*/subdir). (Like arrow, duckdbfs::open_dataset will assume recursive path discovery on directories). Note that http(s) URLs will always require the full vector since a ls() method is not possible. Even with URLs or vector-based paths, duckdb can automatically populate column names given only by hive structure when hive_style=TRUE (default). Note that passing a vector of paths can be significantly faster than globbing with S3 sources where the ls() operation is relatively expensive when there are many partitions.

Performance notes

• In some settings, duckdbfs::open_dataset can give substantially better performance (orders of magnitude) than arrow::open_dataset(), while in other settings it may be comparable or even slower. Package versions, system libraries, network architecture, remote storage performance, network traffic, and other factors can all influence performance, making precise benchmark comparisons in real-world contexts difficult.
• On slow network connections or when accessing a remote table repeatedly, it may improve performance to create a local copy of the table rather than perform all operations over the network. The simplest way to do this is by setting the mode = "TABLE" instead of “VIEW” on open dataset. It is probably desirable to pass a duckdb connection backed by persistent disk location in this case instead of the default cached_connection() unless available RAM is not limiting.
• unify_schema is very computationally expensive. Ensuring all files/partitions match schema in advance or processing different files separately can greatly improve performance.