Data Viewer

Raven includes a data viewer that overrides R's View() so calls in a Raven-managed R terminal open a virtualized grid in a VS Code webview instead of the default backup viewer. The grid streams row windows from disk, keeping scrolling responsive on multi-million-row data frames.

Why we built this

Raven serializes the frame to an Apache Arrow IPC (Feather v2) file and decodes only the rows currently visible, so the webview's memory and decode time scale with the visible viewport rather than the size of the frame. Writing to an on-disk Arrow file also means paging doesn't have to go back through R, which keeps scrolling on multi-million-row frames smooth.

Other R data viewers in VS Code take different staging routes — the REditorSupport extension's sess helper, for example, materializes the whole frame in R memory and serves row windows from there to the webview over JSON-RPC. Both extensions paginate the wire, but the staging is different: Raven snapshots the frame to disk once as Arrow and the webview reads windows from the file directly, bypassing R for paging; sess keeps everything live in R. In our own smoke tests, the on-disk Arrow path has stayed responsive on multi-million-row frames where the R-staged path can hit memory limits. The two viewers also differ on value-labelled data: Raven recognizes haven_labelled plus foreign / readstata13 label maps and substitutes the label when its Labels toggle is on (see Labels); sess classifies labelled columns as formatted numerics and renders the underlying codes. See Comparison: Data viewer for the side-by-side.

What it shows

• data.frame (including tibble and data.table subclasses).

• matrix. If the matrix has non-default rownames, they appear as a leading rowname column; auto-generated 1..N rownames are dropped in favor of the viewer's row-number gutter.

• Atomic vectors and scalars (numeric, integer, character, logical, complex, Date, POSIXct), plus standalone factor and haven_labelled vectors. A vector renders as a values column (a value column when it has length 1). If the vector has names(), those appear in a leading name column — the same role the rowname column plays for a matrix. A standalone factor / haven_labelled vector keeps its labels, so the Labels toggle works on it. A 1-D array (array(1:3)) is rendered the same way as a vector; its dimnames become the name column.

• Flat lists — a list whose elements are all scalars/vectors. Each element becomes a column, padded with empty (NA) cells to the length of the longest element, and each column keeps its own type. Unnamed elements get V1, V2, … headers. A named vector and a list of scalars therefore render differently — c(a = 1, b = 2) is a two-row name/values table, while list(a = 1, b = 2) is a one-row, two-column table — matching as.data.frame(list(a = 1, b = 2)).

• Other objects raise an error in R that explains why:

  > View(sum)
  Error: Can't `View()` an object of class `function`
  

  • A nested/recursive list (an element that is itself a list or data.frame, or otherwise not a plain vector) names the offending element rather than blaming the list class, e.g. Can't `View()` this list: element `b` has class `list`. Only flat lists (every element a vector) are supported. An empty list reports Can't `View()` an empty list.
  • An array with more than two dimensions (a 3-D+ array or a 3-way+ table) reports its dimension count: Can't `View()` an array with 3 dimensions; the data viewer shows 2-dimensional tables only.
  • raw vectors, environments, functions, and other unsupported classes keep the class-based message shown above.

Triggering

Call View() in a Raven-managed R terminal:

View(mtcars)
View(head(iris, 50))
View(my_df, "Custom panel name")

The panel title comes from the second argument when supplied, otherwise from the deparsed expression (truncated to 256 bytes with … if longer). A second View(mtcars) reuses the existing mtcars tab; a different expression opens a new tab.

Toolbar

rows: 12,345 | Sort: mpg▲ cyl▼ ✕ | Filter: cyl ∈ {6,8} ✕ | [Labels] [Format] [3 digits ▾] | [Columns ▾ <7>]

The Sort and Filter strips are hidden entirely when no sort or filter is active.

When the panel is too narrow to fit the chip strips beside the action buttons, the strips drop onto their own second row so Labels / Format / Columns stay reachable. Each strip still scrolls horizontally on its own row when its chips overflow even that full-width row. The wrap decision holds a small hysteresis band so the toolbar doesn't flap between one and two rows as the panel is resized across the boundary.

Each toggle is filled when active; clicking flips it. The Labels and Format buttons (and the digits dropdown) are hidden when no column in the current data set would be affected by them — e.g. an all-integer matrix hides all three, while a frame with only factors hides Format and digits but keeps Labels. The small badge on Columns shows the count of currently hidden columns (absent when none are hidden).

Labels

Defaults to on in new panels. When on, columns with labels render their labels instead of raw values:

• factor columns swap the integer code (1-based, matching as.integer(factor_col) in R) for the level string.
• haven_labelled columns (from haven::read_dta) and other columns with attr(col, "labels") (haven) or attr(col, "value.labels") (foreign::read.dta, readstata13::read.dta13) swap the cell value for the matching label, falling back to the raw value when no label is mapped.
• Columns without label metadata are unchanged.

The variable label (attr(col, "label")) is always shown in the column-header tooltip, regardless of the toggle.

Format

Defaults to on in new panels. When on, non-integer numeric columns are rounded to the digits chosen in the dropdown (0–6, default 3). Integer columns, dates, timestamps, factors, and string columns are unaffected. NaN and ±Inf are rendered as NaN / Inf / -Inf literally — they aren't formatted away.

A Float column that the source file flagged as integer-display (Stata %w.0f, SAS/SPSS F8.0, COMMA10.0, Z3., etc.) is treated like an integer column — Format does nothing to it, and the toggle is hidden when no other column would be affected. Independently, a single integer-valued cell inside an otherwise-decimal Float column (e.g. 5 in a column whose other rows are 1.5, 2.25) is rendered as 5 — not 5.000 — to avoid the misleading trailing-zero look common in SPSS/SAS files that store integer-valued data as doubles.

Columns popover

The Columns ▾ button opens a popover with one checkbox per column. Hide/show changes are persisted per panel-name + schema hash, so the same View(mtcars) opened tomorrow remembers the layout.

Selection and copy

• Click a cell, shift-click another to extend the rectangle.
• Click-drag also extends.
• Click a column header to select that column; click-drag across column headers to select multiple contiguous columns.
• Click a row-number gutter cell to select that row; click-drag across row headers to select multiple contiguous rows.
• Click the # corner cell (top-left) to select every row across all currently-visible columns (a row selection — see the copy note below).
• Cmd/Ctrl+A also selects every row across all currently-visible columns, but as a column selection. The two are visually identical but copy differently: a column selection includes the column-name row, a row selection doesn't.
• Cmd/Ctrl+C copies the selection as TSV. Copying respects the active Labels / Format / digits state — what you see is what you copy.
• Column-header selections and Cmd/Ctrl+A selections include the column-name row when copied. Cell, row-header, and #-corner selections copy data only, matching spreadsheet conventions.
• Right-click a cell, column header, or row-gutter cell to show a Copy menu that copies the current selection. Right-clicking a column header always replaces the selection with that one column (so the menu's Copy copies that column, even if other columns were previously selected). For cell right-clicks, if the target isn't already in the selection it gets selected first; right-clicking a cell inside the existing selection leaves the selection unchanged. The column-header menu also offers Hide column, plus the sort and filter items described under Sorting and Filtering. The platform's default Cut/Copy/Paste menu is suppressed elsewhere in the panel.
• A 5,000,000-cell hard cap protects against accidental huge clipboard writes; over the cap the panel shows a toast and refuses the copy.

Keyboard shortcuts

Key	Action
Home	Jump to the first column of the current row.
End	Jump to the last column of the current row.
Cmd/Ctrl+Home	Jump to the first cell (top-left of the grid).
Cmd/Ctrl+End	Jump to the last cell (bottom-right).
PageUp	Scroll one viewport up.
PageDown	Scroll one viewport down.
Cmd/Ctrl+A	Select all rows across visible columns.
Cmd/Ctrl+C	Copy the current selection as TSV.
Shift+Alt+A	Sort focused column ascending.
Shift+Alt+D	Sort focused column descending.
Shift+Alt+0	Clear all sorts.

These are the data grid's built-in spreadsheet bindings: Home / End move within the current row (first / last column), while Cmd/Ctrl +Home / End jump to the first / last cell of the whole grid.

Sorting

Right-click a column header to sort. The menu offers Sort ascending, Sort descending, and the corresponding Add ascending to sort / Add descending to sort when another column is already sorted, plus Clear sort on this column (only when that column is in the sort) and Clear all sorts (only when some sort is active). Picking Sort replaces the sort with that column; picking Add to sort appends it as the next priority key. Holding Shift when picking Sort ascending / Sort descending is a shortcut for the Add items.

A sorted column shows a hairline triangle on the right edge of its header — ▲ for ascending, ▼ for descending. When more than one column is in the sort, each sorted header also shows a small priority badge (1, 2, 3 …) so you can see which key takes precedence at a glance.

A chip strip appears in the toolbar listing the active keys in priority order. Click any chip to open a small popover with Flip direction, Remove from sort, and (when applicable) Move to first. The trailing ✕ on the strip clears every sort key.

NA / NaN

Missing values — R's NA, floating-point NaN, and NULL — always sort to the bottom in both ascending and descending order, matching order(..., na.last = TRUE). The ±Inf sentinels sort numerically.

Labels and Format

Sort follows what you see in the grid:

• A factor column sorts by integer level when Labels is off and by label string when Labels is on.
• A value-labelled column (haven_labelled, foreign::value.labels, readstata13::read.dta13) sorts by the underlying numeric when Labels is off and by the displayed label (or the raw value when no label exists for a cell) when Labels is on.
• Numeric columns always sort by the underlying double, even when the Format toggle is rounding the display.

Toggling Labels with a sort active on a labelled column will re-sort the rows.

Persistence

Sort state is persisted per panel-name + schema-hash alongside layout and toolbar state, so a later View(df) against the same dataset restores the sort. Only the sort keys (plus a snapshot of the Labels state) are stored — the row permutation is always recomputed against the current data on restore, because schema-hash equality is not evidence that two datasets share row values (column names and types can match while values differ). Set raven.dataViewer.persistSort to false to make every panel open unsorted.

Reopening with a saved sort or filter

The grid itself paints instantly because it is virtualized — only the visible window of rows is read at a time. Restoring a saved sort or filter, however, requires reading the relevant column(s) in full to recompute the permutation / survivor set, which on a multi-million-row frame can take a few seconds. During that time the toolbar shows "Applying your saved sort & filter…" (worded for whichever applies) with a Skip and show data now link below it. The message only appears if the restore takes longer than ~200 ms, so small datasets never flash it.

Skip and show data now abandons the restore, forgets the saved sort/filter for that panel (clears the persisted entries), and shows the data in its natural, unsorted/unfiltered order. This is distinct from a genuine read failure: if reapplying the saved sort/filter fails for a real reason (a corrupt or unreadable column), the panel opens in natural order, a notice explains that the saved sort/filter could not be reapplied, and the saved preferences are kept so the next reopen can retry.

Sort indicator

The active sort keys appear as chips in the toolbar — one per key, in priority order. While the host is building a large permutation, a Sorting… pill appears beside the chips; it is the only progress cue, as the data viewer has no bottom status bar.

Filtering

Right-click a column header and choose Filter… to open the filter editor for that column. You can also press ⇧⌥F to open the editor for the currently focused column.

When the column already has an active filter, the editor opens pre-populated with that filter's current settings — the same predicate, values, and Include NA / NaN state it had when you last applied it — so you edit in place rather than starting over. In that case the context-menu item reads Edit filter…, and applying updates the existing filter (it does not add a second one — see Composition). On an unfiltered column the item reads Filter… and the editor opens empty.

Per-type predicates

Column type	Available predicates
Numeric (Int, Float)	=, ≠, <, ≤, >, ≥, between, not between
Labelled numeric (haven_labelled, foreign, readstata13)	is one of, is not one of (label checklist, matched by code) plus the full numeric set above
Factor	is one of, is not one of
Character	contains, does not contain, starts with, ends with, =, ≠, matches regex
Boolean	is true, is false
Date / Timestamp	=, ≠, <, ≤, >, ≥, between, not between
Any column	is empty, is not empty

For numeric between / not between, the editor shows a histogram of the column's distribution with draggable range thumbs alongside the numeric input fields. The histogram is computed the first time you open that column's filter (not up front when the table loads), so on a very large frame it may appear a moment after the popover opens; it is cached thereafter. This keeps the histogram scan off the initial load path, so opening a table paints the grid without waiting for every numeric column to be scanned.

For factor columns, the editor shows a searchable checklist of levels when the column's value dictionary has been shipped; large or unshipped dictionaries fall back to a free-text value list. For labelled numeric columns the editor defaults to a searchable checklist of the labelled values, each shown as label + underlying code, and also offers the full set of numeric predicates via the condition dropdown.

The matches regex predicate for character columns uses ECMAScript (JavaScript) regex syntax — not PCRE or R regex syntax. A Case sensitive toggle controls whether all character predicates match case-sensitively.

Use is empty to select rows where a column is missing; is not empty selects only non-missing rows.

NA / NaN

By default, missing values (NA, NaN) fail every predicate and are excluded from filtered results. Each filter editor has an Include NA / NaN checkbox that re-includes missing rows for that specific filter.

Labels and Format

Filters on labelled columns match the displayed string — the label when Labels is on, the underlying code or value when Labels is off. This is the same WYSIWYG rule as sorting: what you see is what is matched. Toggling Labels with an active filter on a labelled column will re-evaluate the filter.

Labelled numeric columns are the exception: their is one of / is not one of checklist matches the underlying numeric code, not the displayed label, so toggling Labels never changes which rows a labelled-numeric filter keeps. The chip still shows the labels (for example gender ∈ {Male, Female}). The numeric compare / between predicates likewise match the underlying value, as for any numeric column.

The Format toggle and digits setting never affect filtering. Numeric predicates always match the underlying double, regardless of how many digits are shown in the grid.

Chip strip

Active filters appear as chips in the toolbar. Each chip shows a short summary of the filter condition (for example, cyl ∈ {6, 8} or mpg > 20). Click a chip to re-open its editor. The kebab menu on each chip offers Edit, Enable / Disable (disabled chips are shown greyed-out and are not applied), and Remove. The trailing ✕ on the strip clears all filters.

Composition

Filters across different columns are combined with AND — a row must satisfy all active filters to appear. At most one filter is allowed per column; adding a second filter for the same column replaces the first.

Filter indicator

Active filters appear as chips in the toolbar. The row-count readout in the top-left reflects the filtered total — Showing X-Y of N, where N is the count of rows passing all filters. While the host is rebuilding the filter index a Filtering… pill appears beside the chips.

Persistence

Active filters are persisted per panel-name × schema-hash alongside layout, toolbar state, and sort, so a later View(df) against the same dataset restores the filter configuration. Only the chip descriptors (predicate type, values, enabled state, and the Include NA / NaN flag) are stored — filter membership is always recomputed against the current data on restore. Set raven.dataViewer.persistFilters to false to open every panel unfiltered.

Keyboard shortcuts

Key	Action
Shift+Alt+F	Open filter editor for the focused column.
Shift+Alt+X	Clear filter on the focused column.
Shift+Alt+9	Clear all active filters.
Enter (in editor)	Apply the filter and close the editor.
Escape (in editor)	Cancel without applying.

Settings

Setting	Default	Description
raven.dataViewer.missingValueStyle	foreground	How NA / NaN cells are highlighted: foreground (colorize the text), background (tint the cell), or none.
raven.dataViewer.maxStoredLayouts	10000	LRU cap on persisted column-width / visibility entries. Each unique panel-name × schema-hash pair counts once.
raven.dataViewer.defaultDigits	3	Initial digits used when the Format toggle is on (Format defaults to on). Accepts 0–15; the toolbar's digits dropdown only exposes 0–6, so values above 6 only take effect as the initial digits and can't be picked from the dropdown afterwards.
raven.dataViewer.persistSort	true	Persist the active row sort per panel-name × schema hash. Set to false to make every View(df) open unsorted.
raven.dataViewer.persistFilters	true	Persist active filters per panel-name × schema hash. Set to false to open every panel unfiltered.

The data viewer's overall enable/disable is controlled by raven.rConsole.activation — there is no separate raven.dataViewer.enabled toggle.

Changes apply to newly-opened panels.

How it works

1. The Raven bootstrap profile (loaded via R_PROFILE_USER when you start a Raven-managed R terminal) installs a custom View() in globalenv() before the plot bridge runs. A failure in the plot bridge does not affect View().
2. On View(df), R writes the data frame to an Apache Arrow IPC (Feather v2) file in <extension globalStorage>/data-viewer/. Per-column variable-label and value-label metadata is captured into a single schema-level JSON blob (R's arrow package doesn't expose per-field metadata writes through its public API).
3. R POSTs the path to the extension's loopback HTTP server (POST /view-data with { sessionId, panelName, filePath, nrow }). The route validates that the canonical path is under the per-extension data-viewer directory, rejecting anything else.
4. The extension opens the file via apache-arrow (JS), indexes record-batch starts, and serves row windows on demand to the webview over postMessage. Decoded batches are LRU-cached.

The same Arrow file backs the panel for its lifetime; closing the panel deletes the file. On extension activation the data-viewer directory is swept of files older than 24 hours.

Requirements

• The R arrow package (install.packages("arrow")). If it's missing, Raven prints a warning in R and shows a VS Code warning notification when you call View(), then returns without interrupting the rest of your code.
• A Raven-managed R terminal (the standard "R" terminal profile in Raven, or one launched via Raven's send-to-R commands). Plain R terminals you opened outside Raven won't have the override installed.

Troubleshooting

• Nothing happens when I call View(df). Check that the terminal was started by Raven (the terminal profile dropdown's "R" entry, or via send-to-R). Confirm requireNamespace("arrow") returns TRUE. Check raven.rConsole.activation: if it's "disabled", or "auto" while REditorSupport is enabled or you're in Positron, Raven's R console — and the data viewer — won't activate.
• "Raven data viewer requires the 'arrow' package" warning. Run install.packages("arrow") in the same R installation.
• The Labels toggle doesn't change a column. The column has no label metadata. For haven_labelled columns, this means attr(col, "labels") is empty.
• Copying a huge selection refuses with "Selection exceeds copy limit". The 5 M-cell cap is intentional; reduce the selection or export the slice through R instead.