duckdb

Compatibility

• The dbplyr backend needs dbplyr 2.6.0 or later, and warns when an older dbplyr is loaded (#2605).
• The installed package no longer ships the cpp11 headers under include/ (#2649).
• The package compiles without warnings on the clang 23 and R-devel check flavors (#2617, #2698).
• duckdb() and dbConnect() fail instead of silently ignoring settings a cached database instance cannot apply: a different read_only, a differing config entry, or home, shared_home, allow_extensions or environment_scan. dbConnect(drv, dbdir) also fails when drv already owns a different database file. Call duckdb_shutdown() first, or pass the settings to the duckdb() call that creates the instance (#2641).

Bug fixes

Crashes and hangs

• A data frame parameter binds one struct per row instead of crashing R (#2818).
• A list column of data frames or factors no longer crashes or hangs a parallel scan (#2834), and lazy (ALTREP) vectors inside struct, matrix and list columns are materialized on R's thread before a scan reads them (#2582).
• Writing a named list with an empty value under map = "list_of" no longer crashes (#2866).
• Converting VARIANT values to R no longer exposes intermediate objects to the garbage collector (#2750, #2753).
• The progress display keeps its R callback alive and calls it only on R's thread (#2828), and prints at most every half second (#2844).
• Opening a database whose previous instance is still held by a connection or an unfinished result no longer spins at full CPU, it reports that the database is still in use (#2759).
• A failure inside a lazy data frame's ALTREP method no longer strips duckdb_error from later errors or leaves a partly converted column behind (#2829), and an error raised while such a method runs no longer calls into R (#1796, #1797).

Types and time zones

• TIMESTAMPTZ columns take their tzone attribute from the session's TimeZone setting, also in lazy relations, and stay in UTC when the icu extension is not available (#184, #2401).
• dbQuoteLiteral() keeps the sub-second part of a POSIXct value (#2646, #2763).
• integer64 columns and parameters are always written as BIGINT, whatever the bigint setting (#2819).
• Integer-backed Date and difftime columns keep their NA values, and a data frame column with several fields writes as STRUCT (#2808).
• A table whose columns R cannot hold is still found and listed, and a query returning such a column fails before it runs, naming the column (#2808).

Connections and storage

• An in-memory database spills to its own directory below tempdir() out of the box, created only when a query spills and removed when the instance shuts down. An on-disk database keeps DuckDB's <dbdir>.tmp default, as the DuckDB CLI does (#2562).
• ~/.duckdb is no longer created for a session that only looks interactive, such as one with options(rlang_interactive = TRUE). Without a person to answer the prompt, the per-session directory is used (#2550).
• A database path is normalized only as far as it resolves, so a path that cannot be resolved is handed to the engine rather than refused (#455, #2623), and a dbdir that cannot be created is reported against the user's call (#2822).
• A dbdir that an extension handles, such as a URL, is no longer treated as a file (#2641).
• dbIsValid() on a driver no longer opens the database (#2754).

Queries and results

• A PRAGMA runs after the statements before it in one query string (@t-kalinowski, #2792). Its expansion no longer leaves a transaction open or refuses a LOAD halfway, and an empty one can end a query (#2845).
• dbListFields() and dbAppendTable() read the columns of a table named by Id() correctly (#2848).
• dbClearResult() frees the unread results of a multi-row bind at once, not at the next garbage collection (#2784, #2825), and ends the query of an Arrow stream that was not read to the end (#2793).

Arrow

• Reading an Arrow stream that another statement invalidated is an error, not the end of the stream (#2775), and an Arrow stream keeps its connection's client context alive (#2790).
• dbFetchArrowChunk() builds the final empty chunk for LIST and INTERVAL columns (#2774), and each dbFetchArrowChunk() and dbFetchArrow() call uses its own chunk_size (#2791).
• duckdb_fetch_arrow() and duckdb_fetch_record_batch() refuse a cleared result, as the other result methods do (#2785, #2821).

Performance

• Release the ALTREP query result after materialization and transformation (@toppyy, #1027).
• The list of valid time zones is read once per session, and again when TZDIR changes (#2786, #2810).

Documentation

• New reference pages ?duckdb_types and ?duckdb_types_arrow catalog every DuckDB type in and out of R, through R vectors and through Arrow, geometry included (#2836, #2837, #2850, #2861).
• The new reference page ?duckdb_memory and the help pages for fetching and writing data describe the memory a query takes (#2666).
• ?duckdb describes what keeps a database instance open and the limitations known today (#2823, #2849).
• The connection's help page describes how a string holding several statements runs (#2835).
• The README states the package's goals and non-goals (#2731).