Get started with datasetviewer

Looking at a dataset should not mean choosing between fast and complete. Grids that render every cell into the page bog down at a few thousand rows; ones that page through a server add a network round-trip to every scroll and filter. The usual escape hatch — show the first 1,000 rows — quietly hides exactly the rows you opened the viewer to find.

datasetviewer takes a different route, borrowed from SAS Studio’s table viewer. The data is sent to the browser once, as Parquet, and queried in place with DuckDB-WASM; the grid is drawn on an HTML canvas and only ever materialises the rows you can see. Sort, filter, hide a column, jump to the last page — each is a SQL query over the whole dataset that returns in milliseconds, with no row sampling. The same widget runs in an interactive Shiny app and in a static HTML document like this one.

Your first viewer

Hand dataset_viewer() a data frame. That is the entire API for the common case — everything else is interaction inside the widget.

library(datasetviewer)
dataset_viewer(mtcars)

Try it: drag the scrollbar, drag a column border to resize, or click a row. The grid above is live — it is the real widget, not a screenshot.

A guided tour of the interface

The layout mirrors SAS Studio, so anyone who has used that viewer is already at home:

• Columns panel (left) — a checklist of every column with a type chip (A for character, # for numeric, a calendar for dates). Uncheck a column to hide it from the grid; the data is never reloaded. Sort the list by original order, name, or type, and filter it by name to find columns in a wide dataset (the list order is a navigation aid; the grid column order is unchanged).
• Property pane (lower left) — select a column to inspect its Label, Name, Length, Type, and Format, the same attributes PROC CONTENTS reports.
• Toolbar (top) — the names-versus-labels View dropdown, an Export current view to CSV button, a Show code button (<>) that reveals the dplyr pipeline for the current view, and Filter table rows (the funnel) with a badge showing the active filter.
• Header sort — click a column header to select it, then click again to cycle its sort: ascending, descending, and back to unsorted. Shift-click further headers to build a multi-column sort; each sorted column shows its direction and priority (AGE ↑1, SEX ↓2).
• Header menu — right-click any column header to sort it (Sort Ascending / Descending add the column to the sort; Clear Sorting removes just that column), add a filter, copy the column or its header, or size the columns to content.
• Status bar — the total row and column counts, and the filtered count once a filter is active.

CDISC metadata, labels, and the property pane

A plain data frame has no labels, so the property pane shows names only. Point the viewer at a labelled or CDISC-conformed frame and the metadata comes to life. With the companion artoo package installed, column labels, formats, and storage lengths are read straight from the frame and shown in the property pane — and you can set the header row to use labels instead of names.

# artoo ships the CDISC pilot ADaM datasets used across these docs.
dataset_viewer(artoo::cdisc_adsl, view = "labels")

Select STUDYID in the columns panel and the property pane reads Study Identifier; the header row now shows labels because of view = "labels". No artoo dependency is required for plain frames — it is consulted only when present.

Filtering the whole table

There are two ways to filter, both operating over every row:

1. Filter Table Rows — click the funnel in the toolbar and type a free-text expression, SAS-style, such as AGE >= 75 and SEX = "F". It is translated to a SQL WHERE clause, and the status bar updates to the matched count.
2. Add Filter — right-click a column header. The dialog adapts to the column’s type: a checklist of distinct values for character columns, a comparison operator and value for numbers, and a date picker for dates.

Because the filter runs in DuckDB over the full Parquet payload, the answer is exact — the matched count is the true count, not a count within a sampled window.

Reproducing the view as code

Exploration in the grid is convenient, but a report needs to be reproducible. The Show code button (<> in the toolbar) opens a dialog with the runnable dplyr pipeline that reproduces the current view — the filter, the sort, and the column selection, in order:

library(dplyr)

mtcars |>
  filter(mpg >= 20) |>
  arrange(desc(hp)) |>
  select(cyl, hp, wt, mpg)

select() comes last so the filter and the sort can reference a column the view hides — narrowing first would drop it before those steps run.

The snippet is air-formatted and syntax-highlighted, with a Copy button. SQL idioms are translated to their R equivalents — IN (...) becomes %in% c(...), NOT IN becomes !x %in% c(...), and date or time literals become as.Date() / as.POSIXct() / hms::as_hms() — so the code runs as-is against the source frame. It is modelled on SAS Studio’s “show the code that creates this table”, and it stays in sync with the view: change the filter or sort and reopen it to see the updated pipeline.

Exporting the current view

The Export current view to CSV toolbar button downloads exactly what you are looking at — the visible columns, the active filter, and the current sort, over every matching row, not just the visible window. The export streams from the engine in row chunks, so it does not depend on the dataset fitting in memory in one piece.

Built for scale: no row sampling

The design choice that makes this work is moving the query engine into the browser:

• Transport. The frame is serialised to Parquet with nanoparquet and carried in the widget payload — columnar, compressed, and read natively by the engine.
• Engine. DuckDB-WASM reads that Parquet directly and answers every filter, sort, and page as SQL.
• Grid. The canvas grid asks the engine only for the rows in the visible window (a LIMIT/OFFSET query), so scrolling cost is independent of the dataset’s size.

The practical upshot: a viewer over a multi-million-row frame scrolls, sorts, and filters as smoothly as one over mtcars, and every row stays reachable.

Embedding in a Shiny app

In Shiny, pair datasetviewerOutput() in the UI with renderDatasetViewer() on the server. The viewer is not a dead end: the user’s current column selection, filter, sort, and view mode flow back into the app as inputs, namespaced by the output id, so the rest of the app can react to what the analyst is looking at.

library(shiny)
library(datasetviewer)

ui <- fluidPage(
  datasetviewerOutput("viewer", height = "560px"),
  verbatimTextOutput("state")
)

server <- function(input, output, session) {
  output$viewer <- renderDatasetViewer(dataset_viewer(mtcars))

  # State changes in the widget arrive as inputs, namespaced by output id.
  output$state <- renderPrint({
    list(
      columns = input$viewer_columns, # columns currently shown
      filter  = input$viewer_filter, # active filter expression
      sort    = input$viewer_sort, # active sort
      view    = input$viewer_view # "names" or "labels"
    )
  })
}

shinyApp(ui, server)

Static HTML and Quarto

No server is required for the static case — this very vignette embeds live widgets. Drop dataset_viewer() into any R Markdown or Quarto document and the result is a fully interactive grid. The same call you would write in a Shiny app produces the same viewer here.

The query engine: online by default, offline when you need it

Everything except one piece is bundled in the package and works with no internet: the canvas grid, the column panel, the filters, the code view, the CSV export, and your data (carried in the page as Parquet). The one piece is the DuckDB-WASM query engine — the in-browser database that answers every filter, sort, and page. It is roughly 35 MB, far too large to ship inside an R package, so by default the widget loads it from a public CDN (jsDelivr) the first time a grid is rendered. For interactive use on a connected machine, nothing more is needed.

Self-hosting the engine for offline and corporate use

When the browser cannot reach the CDN — an air-gapped laptop, or a corporate Shiny server behind a firewall — the engine must be served locally. This works the same way the arrow package acquires its C++ library: at install time, with no function for you to call.

When you install datasetviewer, an install step fetches the engine (and the parquet extension DuckDB needs to read the payload) into the package. From then on, a Shiny app serves the engine from the package to the browser — no internet at runtime. If the install machine cannot reach the public host, the step is skipped and the widget simply falls back to the CDN; the install never fails.

The fetch is steered with environment variables, set before install.packages("datasetviewer") (the analogues of arrow’s LIBARROW_BINARY):

A typical corporate deployment installs the package the same way it installs any other (often through an internal mirror that already carries arrow), points these variables at the in-house mirror if the public host is blocked, and then runs the Shiny app — which serves the engine to every user’s browser with no outbound connection.

Static documents and the engine

A self-contained HTML document (a Quarto or R Markdown report, or htmlwidgets::saveWidget(selfcontained = TRUE)) embeds every dependency in the file. Embedding the 35 MB engine there would produce an enormous page, so static documents should load the engine from the CDN instead. Set this option once, in a setup chunk, before any dataset_viewer() call:

options(datasetviewer.use_local_engine = FALSE)

That keeps the document small while the grid loads the engine from the CDN when a reader opens it. (This vignette does exactly that.) Leave the option at its default for Shiny, where the engine is served rather than embedded.

Where to next

• ?dataset_viewer — the full argument reference, including view, width, and height.
• ?datasetviewerOutput and ?renderDatasetViewer — the Shiny bindings and the input names the widget publishes.
• artoo — lossless CDISC dataset I/O and the metadata model the property pane reads.