• pivot_wider now uses .by and |> syntax for the dplyr helper message to
  identify duplicates (@boshek, #1516)

New features

• New family of consistent string separating functions:
  separate_wider_delim(), separate_wider_position(),
  separate_wider_regex(), separate_longer_delim(), and
  separate_longer_position(). These functions are thorough refreshes of
  separate() and extract(), featuring improved performance, greater
  consistency, a polished API, and a new approach for handling problems. They
  use stringr and supersede extract(), separate(), and separate_rows()
  (#1304).

• nest() gains a .by argument which allows you to specify the columns to
  nest by (rather than the columns to nest, i.e. through ...). Additionally,
  the .key argument is no longer deprecated, and is used whenever ... isn't
  specified (#1458).

• unnest_longer() gains a keep_empty argument like unnest() (#1339).

• pivot_longer() gains a cols_vary argument for controlling the ordering of
  the output rows relative to their original row number (#1312).

• New datasets who2, household, cms_patient_experience, and
  cms_patient_care to demonstrate various tidying challenges (#1333).

Breaking changes

• The ... argument of both pivot_longer() and pivot_wider() has been
  moved to the front of the function signature, after the required arguments
  but before the optional ones. Additionally, pivot_longer_spec(),
  pivot_wider_spec(), build_longer_spec(), and build_wider_spec() have
  all gained ... arguments in a similar location. This change allows us to
  more easily add new features to the pivoting functions without breaking
  existing CRAN packages and user scripts.

  pivot_wider() provides temporary backwards compatible support for the case
  of a single unnamed argument that previously was being positionally matched to
  id_cols. This one special case still works, but will throw a warning
  encouraging you to explicitly name the id_cols argument.

  To read more about this pattern, see
  Data, dots, details in the
  tidyverse design guide (#1350).

Lifecycle changes

• All functions deprecated in tidyr 1.0 and 1.2 (the old lazyeval functions
  ending in _ and various arguments to unnest()) now warn on every use.
  They will be made defunct in 2024 (#1406).

Rectangling

• unnest_longer() now consistently drops rows with either NULL or empty
  vectors (like integer()) by default. Set the new keep_empty argument to
  TRUE to retain them. Previously, keep_empty = TRUE was implicitly being
  used for NULL, while keep_empty = FALSE was being used for empty vectors,
  which was inconsistent with all other tidyr verbs with this argument (#1363).

• unnest_longer() now uses "" in the index column for fully unnamed
  vectors. It also now consistently uses NA in the index column for empty
  vectors that are "kept" by keep_empty = TRUE (#1442).

• unnest_wider() now errors if any values being unnested are unnamed and
  names_sep is not provided (#1367).

• unnest_wider() now generates automatic names for partially unnamed
  vectors. Previously it only generated them for fully unnamed vectors,
  resulting in a strange mix of automatic names and name-repaired names (#1367).

Bug fixes and minor improvements

General

• Most tidyr functions now consistently disallow renaming during tidy-selection.
  Renaming was never meaningful in these functions, and previously either had no
  effect or caused problems (#1449, #1104).

• tidyr errors (including input validation) have been thoroughly reviewed
  and should generally be more likely to point you in the right direction
  (#1313, #1400).

• uncount() is now generic so implementations can be provided for objects
  other than data frames (@mgirlich, #1358).

• uncount() gains a ... argument. It comes between the required and the
  optional arguments (@mgirlich, #1358).

• nest(), complete(), expand(), and fill() now document their support
  for grouped data frames created by dplyr::group_by() (#952).

• All built in datasets are now standard tibbles (#1459).

• R >=3.4.0 is now required, in line with the tidyverse standard of supporting
  the previous 5 minor releases of R.

• rlang >=1.0.4 and vctrs >=0.5.2 are now required (#1344, #1470).

• Removed dependency on ellipsis in favor of equivalent functions in rlang
  (#1314).

Nesting, packing, and chopping

• unnest(), unchop(), unnest_longer(), and unnest_wider() better handle
  lists with additional classes (#1327).

• pack(), unpack(), chop(), and unchop() all gain an error_call
  argument, which in turn improves some of the error calls shown in nest()
  and various unnest() adjacent functions (#1446).

• chop(), unpack(), and unchop() all gain ..., which must be empty
  (#1447).

• unpack() does a better job of reporting column name duplication issues and
  gives better advice about how to resolve them using names_sep. This also
  improves errors from functions that use unpack(), like unnest() and
  unnest_wider() (#1425, #1367).

Pivoting

• pivot_longer() no longer supports interpreting values_ptypes = list()
  and names_ptypes = list() as NULL. An empty list() is now interpreted as
  a <list> prototype to apply to all columns, which is consistent with how any
  other 0-length value is interpreted (#1296).

• pivot_longer(values_drop_na = TRUE) is faster when there aren't any missing
  values to drop (#1392, @mgirlich).

• pivot_longer() is now more memory efficient due to the usage of
  vctrs::vec_interleave() (#1310, @mgirlich).

• pivot_longer() now throws a slightly better error message when
  values_ptypes or names_ptypes is provided and the coercion can't be made
  (#1364).

• pivot_wider() now throws a better error message when a column selected by
  names_from or values_from is also selected by id_cols (#1318).

• pivot_wider() is now faster when names_sep is provided (@mgirlich, #1426).

• pivot_longer_spec(), pivot_wider_spec(), build_longer_spec(), and
  build_wider_spec() all gain an error_call argument, resulting in better
  error reporting in pivot_longer() and pivot_wider() (#1408).

Missing values

• fill() now works correctly when there is a column named .direction in
  data (#1319, @tjmahr).

• replace_na() is faster when there aren't any missing values to replace
  (#1392, @mgirlich).

• The documentation of the replace argument of replace_na() now mentions
  that replace is always cast to the type of data (#1317).

• Hot patch release to resolve R CMD check failures.

Pivoting

• pivot_wider() gains new names_expand and id_expand arguments for turning
  implicit missing factor levels and variable combinations into explicit ones.
  This is similar to the drop argument from spread() (#770).

• pivot_wider() gains a new names_vary argument for controlling the
  ordering when combining names_from values with values_from column names
  (#839).

• pivot_wider() gains a new unused_fn argument for controlling how to
  summarize unused columns that aren't involved in the pivoting process (#990,
  thanks to @mgirlich for an initial implementation).

• pivot_longer()'s names_transform and values_transform arguments now
  accept a single function which will be applied to all of the columns
  (#1284, thanks to @smingerson for an initial implementation).

• pivot_longer()'s names_ptypes and values_ptypes arguments now
  accept a single empty ptype which will be applied to all of the columns
  (#1284).

Nesting

• unnest() and unchop()'s ptype argument now accepts a single empty
  ptype which will be applied to all cols (#1284).

• unpack() now silently skips over any non-data frame columns specified by
  cols. This matches the existing behavior of unchop() and unnest()
  (#1153).

Rectangling

• unnest_wider() and unnest_longer() can now unnest multiple columns at
  once (#740).

• unnest_longer()'s indices_to and values_to arguments now accept
  a glue specification, which is useful when unnesting multiple columns.

• For hoist(), unnest_longer(), and unnest_wider(), if a ptype is
  supplied, but that column can't be simplified, the result will be a list-of
  column where each element has type ptype (#998).

• unnest_wider() gains a new strict argument which controls whether or not
  strict vctrs typing rules should be applied. It defaults to FALSE for
  backwards compatibility, and because it is often more useful to be lax
  when unnesting JSON, which doesn't always map one-to-one with R's types
  (#1125).

• hoist(), unnest_longer(), and unnest_wider()'s simplify argument now
  accepts a named list of TRUE or FALSE to control simplification on a per
  column basis (#995).

• hoist(), unnest_longer(), and unnest_wider()'s transform argument now
  accepts a single function which will be applied to all components (#1284).

• hoist(), unnest_longer(), and unnest_wider()'s ptype argument now
  accepts a single empty ptype which will be applied to all components (#1284).

Grids

• complete() gains a new explicit argument for limiting fill to only
  implicit missing values. This is useful if you don't want to fill in
  pre-existing missing values (#1270).

• complete() gains a grouped data frame method. This generates a more correct
  completed data frame when groups are involved (#396, #966).

• complete() and expand() no longer allow you to complete or expand on a
  grouping column. This was never well-defined since completion/expansion on a
  grouped data frame happens "within" each group and otherwise has the
  potential to produce erroneous results (#1299).

Missing values

• drop_na(), replace_na(), and fill() have been updated to utilize vctrs.
  This means that you can use these functions on a wider variety of column
  types, including lubridate's Period types (#1094), data frame columns, and
  the rcrd type from vctrs.

• replace_na() no longer allows the type of data to change when the
  replacement is applied. replace will now always be cast to the type of
  data before the replacement is made. For example, this means that using a
  replacement value of 1.5 on an integer column is no longer allowed.
  Similarly, replacing missing values in a list-column must now be done with
  list("foo") rather than just "foo".

• replace_na() no longer replaces empty atomic elements in list-columns
  (like integer(0)). The only value that is replaced in a list-column is
  NULL (#1168).

• drop_na() no longer drops empty atomic elements from list-columns
  (like integer(0)). The only value that is dropped in a list-column is
  NULL (#1228).

Bug fixes and minor improvements

General

• @mgirlich is now a tidyr author in recognition of his significant and
  sustained contributions.

• All lazyeval variants of tidyr verbs have been soft-deprecated. Expect them
  to move to the defunct stage in the next minor release of tidyr (#1294).

• any_of() and all_of() from tidyselect are now re-exported (#1217).

• dplyr >= 1.0.0 is now required.

Pivoting

• pivot_wider() now gives better advice about how to identify duplicates when
  values are not uniquely identified (#1113).

• pivot_wider() now throws a more informative error when values_fn doesn't
  result in a single summary value (#1238).

• pivot_wider() and pivot_longer() now generate more informative
  errors related to name repair (#987).

• pivot_wider() now works correctly when values_fill is a data frame.

• pivot_wider() no longer accidentally retains values_from when pivoting
  a zero row data frame (#1249).

• pivot_wider() now correctly handles the case where an id column name
  collides with a value from names_from (#1107).

• pivot_wider() and pivot_longer() now both check that the spec columns
  .name and .value are character vectors. Additionally, the .name
  column must be unique (#1107).

• pivot_wider()'s names_from and values_from arguments are now
  required if their default values of name and value don't correspond to
  columns in data. Additionally, they must identify at least 1 column
  in data (#1240).

• pivot_wider()'s values_fn argument now correctly allows anonymous
  functions (#1114).

• pivot_wider_spec() now works correctly with a 0-row data frame and a spec
  that doesn't identify any rows (#1250, #1252).

• pivot_longer()'s names_ptypes argument is now applied after
  names_transform for consistency with the rectangling functions
  (i.e. hoist()) (#1233).

• check_pivot_spec() is a new developer facing function for validating a pivot
  spec argument. This is only useful if you are extending pivot_longer() or
  pivot_wider() with new S3 methods (#1087).

Nesting

• The nest() generic now avoids computing on .data, making it more
  compatible with lazy tibbles (#1134).

• The .names_sep argument of the data.frame method for nest() is now
  actually used (#1174).

• unnest()'s ptype argument now works as expected (#1158).

• unpack() no longer drops empty columns specified through cols (#1191).

• unpack() now works correctly with data frame columns containing 1 row but
  0 columns (#1189).

• chop() now works correctly with data frames with 0 rows (#1206).

• chop()'s cols argument is no longer optional. This matches the
  behavior of cols seen elsewhere in tidyr (#1205).

• unchop() now respects ptype when unnesting a non-list column (#1211).

Rectangling

• hoist() no longer accidentally removes elements that have duplicated names
  (#1259).

Grids

• The grouped data frame methods for complete() and expand() now move the
  group columns to the front of the result (in addition to the columns you
  completed on or expanded, which were already moved to the front). This should
  make more intuitive sense, as you are completing or expanding "within" each
  group, so the group columns should be the first thing you see (#1289).

• complete() now applies fill even when no columns to complete are
  specified (#1272).

• expand(), crossing(), and nesting() now correctly retain NA values of
  factors (#1275).

• expand_grid(), expand(), nesting(), and crossing() now silently
  apply name repair to automatically named inputs. This avoids a number of
  issues resulting from duplicate truncated names
  (#1116, #1221, #1092, #1037, #992).

• expand_grid(), expand(), nesting(), and crossing() now allow
  columns from unnamed data frames to be used in expressions after that
  data frame was specified, like expand_grid(tibble(x = 1), y = x). This
  is more consistent with how tibble() behaves.

• expand_grid(), expand(), nesting(), and crossing() now work
  correctly with data frames containing 0 columns but >0 rows (#1189).

• expand_grid(), expand(), nesting(), and crossing() now return a 1
  row data frame when no inputs are supplied, which is more consistent with
  prod() == 1L and the idea that computations involving the number of
  combinations computed from an empty set should return 1 (#1258).

Missing values

• drop_na() no longer drops missing values from all columns when a tidyselect
  expression that results in 0 columns being selected is used (#1227).

• fill() now treats NaN like any other missing value (#982).

• expand_grid() is now about twice as fast and pivot_wider() is a bit faster
  (@mgirlich, #1130).

• unchop() is now much faster, which propagates through to various functions,
  such as unnest(), unnest_longer(), unnest_wider(), and
  separate_rows() (@mgirlich, @DavisVaughan, #1127).

• unnest() is now much faster (@mgirlich, @DavisVaughan, #1127).

• unnest() no longer allows unnesting a list-col containing a mix of vector
  and data frame elements. Previously, this only worked by accident, and is
  considered an off-label usage of unnest() that has now become an error.

• tidyr verbs no longer have "default" methods for lazyeval fallbacks. This
  means that you'll get clearer error messages (#1036).

• uncount() error for non-integer weights and gives a clearer error message
  for negative weights (@mgirlich, #1069).

• You can once again unnest dates (#1021, #1089).

• pivot_wider() works with data.table and empty key variables (@mgirlich, #1066).

• separate_rows() works for factor columns (@mgirlich, #1058).

• separate_rows() returns to 1.1.0 behaviour for empty strings
  (@rjpatm, #1014).

• New tidyr logo!

• stringi dependency has been removed; this was a substantial dependency that
  make tidyr hard to compile in resource constrained environments
  (@rjpat, #936).

• Replace Rcpp with cpp11. See https://cpp11.r-lib.org/articles/motivations.html
  for reasons why.

General features

• pivot_longer(), hoist(), unnest_wider(), and unnest_longer() gain
  new transform arguments; these allow you to transform values "in flight".
  They are partly needed because vctrs coercion rules have become stricter,
  but they give you greater flexibility than was available previously (#921).

• Arguments that use tidy selection syntax are now clearly documented and
  have been updated to use tidyselect 1.1.0 (#872).

Pivoting improvements

• Both pivot_wider() and pivot_longer() are considerably more performant,
  thanks largely to improvements in the underlying vctrs code
  (#790, @DavisVaughan).

• pivot_longer() now supports names_to = character() which prevents the
  name column from being created (#961).

  df <- tibble(id = 1:3, x_1 = 1:3, x_2 = 4:6)  
  df %>% pivot_longer(-id, names_to = character())
  

• pivot_longer() no longer creates a .copy variable in the presence of
  duplicate column names. This makes it more consistent with the handling
  of non-unique specs.

• pivot_longer() automatically disambiguates non-unique ouputs, which can
  occur when the input variables include some additional component that you
  don't care about and want to discard (#792, #793).

  df <- tibble(id = 1:3, x_1 = 1:3, x_2 = 4:6)  
  df %>% pivot_longer(-id, names_pattern = "(.)_.")
  df %>% pivot_longer(-id, names_sep = "_", names_to = c("name", NA))
  df %>% pivot_longer(-id, names_sep = "_", names_to = c(".value", NA))
  

• pivot_wider() gains a names_sort argument which allows you to sort
  column names in order. The default, FALSE, orders columms by their
  first appearance (#839). In a future version, I'll consider changing the
  default to TRUE.

• pivot_wider() gains a names_glue argument that allows you to construct
  output column names with a glue specification.

• pivot_wider() arguments values_fn and values_fill can now be single
  values; you now only need to use a named list if you want to use different
  values for different value columns (#739, #746). They also get improved
  errors if they're not of the expected type.

Rectangling

• hoist() now automatically names pluckers that are a single string (#837).
  It error if you use duplicated column names (@mgirlich, #834), and now uses
  rlang::list2() behind the scenes (which means that you can now use !!!
  and :=) (#801).

• unnest_longer(), unnest_wider(), and hoist() do a better job
  simplifying list-cols. They no longer add unneeded unspecified() when
  the result is still a list (#806), and work when the list contains
  non-vectors (#810, #848).

• unnest_wider(names_sep = "") now provides default names for unnamed inputs,
  suppressing the many previous name repair messages (#742).

Nesting

• pack() and nest() gains a .names_sep argument allows you to strip outer
  names from inner names, in symmetrical way to how the same argument to
  unpack() and unnest() combines inner and outer names (#795, #797).

• unnest_wider() and unnest_longer() can now unnest list_of columns. This
  is important for unnesting columns created from nest() and with
  pivot_wider(), which will create list_of columns if the id columns are
  non-unique (#741).

Bug fixes and minor improvements

• chop() now creates list-columns of class vctrs::list_of(). This helps
  keep track of the type in case the chopped data frame is empty, allowing
  unchop() to reconstitute a data frame with the correct number and types
  of column even when there are no observations.

• drop_na() now preserves attributes of unclassed vectors (#905).

• expand(), expand_grid(), crossing(), and nesting() once again
  evaluate their inputs iteratively, so you can refer to freshly created
  columns, e.g. crossing(x = seq(-2, 2), y = x) (#820).

• expand(), expand_grid(), crossing(), and nesting() gain a
  .name_repair giving you control over their name repair strategy
  (@jeffreypullin, #798).

• extract() lets you use NA in into, as documented (#793).

• extract(), separate(), hoist(), unnest_longer(), and unnest_wider()
  give a better error message if col is missing (#805).

• pack()'s first argument is now .data instead of data (#759).

• pivot_longer() now errors if values_to is not a length-1 character vector
  (#949).

• pivot_longer() and pivot_wider() are now generic so implementations
  can be provided for objects other than data frames (#800).

• pivot_wider() can now pivot data frame columns (#926)

• unite(na.rm = TRUE) now works for all types of variable, not just character
  vectors (#765).

• unnest_wider() gives a better error message if you attempt to unnest
  multiple columns (#740).

• unnest_auto() works when the input data contains a column called col
  (#959).

Compatibility with vctrs 0.3.0. Note that because of stricter semantics in vctrs, the names_ptypes and values_ptypes arguments have become less useful. We will fix this in the next version with more general transformation arguments.