# clock 0.7.4

* Avoid non-API `SET_ATTRIB()`.

# clock 0.7.3

* `%Y`, `%F`, `%G`, and `%c` now parse up to 4 _characters_ by default, rather
  than 4 _digits_. This is more in line with the C++20 specification and was a
  change made in the upstream `<date.h>` parser. Practically, this means that
  negative years such as `-1000-01-01` will no longer parse with `%Y-%m-%d`, and
  instead requires `%5Y-%m-%d` to capture the leading `-` (#387).

* tzdb >=0.5.0 is required. Note that older versions of clock are not compatible
  with tzdb 0.5.0 and above, so if you are having issues (particularly with
  parsing) please make sure that both clock and tzdb are fully updated (#387).

* Fixed `-Wdeprecated-literal-operator` reported by clang (#386,
  @MichaelChirico).

* R >=4.0.0 is now required. This is consistent with the standards of the
  tidyverse.

# clock 0.7.2

* Added a `diff()` method for time points and calendars to ensure that durations
  are always returned, even in the empty result case (#364).

* Fixed an issue where clock would not compile on Centos 7 using gcc-5.4.0 due
  to a `constexpr` issue (#357).

* Fixed a test that failed due to `seq.Date()` now returning integer storage in
  some cases in the development version of R.

# clock 0.7.1

* Removed usage of non-API `STRING_PTR()` in favor of `STRING_PTR_RO()`.

* Fixed a gcc warning reported by CRAN related to templated C++ constructors
  (#371).

# clock 0.7.0

## New features

* New `year_week_day()` calendar for specifying a date using the year, the week
  number, and the day of the week, alongside a `start` value representing the
  day of the week that is considered the start of the week. Using
  `start = clock_weekdays$monday` is identical to the `iso_year_week_day()`
  calendar, and using `start = clock_weekdays$sunday` is useful for representing
  the Epidemiological calendar used by the US CDC guidelines (similar to what
  is supported by `lubridate::epiweek()` and `lubridate::epiyear()`) (#110).

* New `date_spanning_seq()` for generating a regular sequence along the full
  span of a date or date-time vector (i.e. along `[min(x), max(x)]`). It is
  similar to `tidyr::full_seq()`, but is a bit simpler and currently has better
  handling of some edge cases. Additionally included in the low-level API are
  `calendar_spanning_seq()`, `time_point_spanning_seq()`, and
  `duration_spanning_seq()` (#279).

* New `date_time_info()` and `zoned_time_info()` low-level helpers for accessing
  the previous/next transition times, the offset from UTC, and the current time
  zone abbreviation (#295).

* `calendar_leap_year()` now supports the year-quarter-day and iso-year-week-day
  calendars (#332, #333).

## Breaking changes

* The storage mechanism for the duration, sys-time, naive-time, and zoned-time
  types has been altered to more correctly represent the full range of values
  allowed by the underlying C++ types. This means that if you have serialized
  a value of one of these types with an old version of clock, then it will no
  longer unserialize correctly going forward.

  Technically, rather than storing a variable number of integer vectors
  representing ticks, ticks of a day, and ticks of a second, we now always store
  values of these types within two double vectors, regardless of the precision.
  This simplifies the implementation and allows us to represent the full range
  of possible `int64_t` values (#331).

## Lifecycle changes

* `date_zone()` and `date_set_zone()` have been soft-deprecated in favor of
  `date_time_zone()` and `date_time_set_zone()` (#326).

## Minor changes and bug fixes

* clock now compiles significantly faster (on a 2018 Intel Mac, it used to take
  ~70 seconds for a full compilation, and now takes ~25 seconds) (#322).

* `%%` and `%/%` operators now return a missing value when the right-hand side
  is `0`. For `%/%`, this is consistent with `2L %/% 0L`, which returns a
  missing value, rather than with `2 %/% 0`, which returns `Inf`, since
  infinite durations are not supported (#349).

* `seq()` methods for durations and time points handle the empty sequence cases
  of `from > to && by > 0` and `from < to && by < 0` better when `from` and `to`
  are very far apart (i.e. when they would otherwise result in overflow if they
  were subtracted).

* `zoned_time_zone()` and `zoned_time_set_zone()` are no longer generic, and now
  only work for zoned-times.

* Documented clock's current stance on leap seconds in the FAQ vignette (clock
  ignores them like POSIXct) (#309).

* Duration vectors now work as `.before` and `.after` arguments of
  `slider::slide_index()` and friends (#306).

* All `as_*()` generics exported by clock now include `...` in their signature
  to help with extensibility of converting to clock types. These are the only
  clock generics that are currently "blessed" as fully extensible (#348).

* `as.character()` has been implemented for durations.

* Fixed `vec_ptype_full()` and `vec_ptype_abbr()` methods for sys-time and
  naive-time objects (#302).

* Many errors have been improved (#219, #286, #595).

* Renamed `locale.h` to `fill.h` to avoid clock's `locale.h` being chosen over a
  system header of the same name on some CentOS machines (#310).

* Skipped a test on 32-bit architectures to work around a bug in base R (#312).

* R >=3.5.0 is now required, which is in line with tidyverse standards.

* vctrs >=0.6.1 and rlang >=1.1.0 are now required.

# clock 0.6.1

* `date_seq()` and the `seq()` methods for the calendar, time point, and
  duration types now allow `from > to` when `by > 0`. This now results in a
  size zero result rather than an error, which is more in line with
  `rlang::seq2()` and generally has more useful programmatic properties (#282).

* The sys-time method for `as.POSIXct()` now correctly promotes to a precision
  of at least seconds before attempting the conversion. This matches the
  behavior of the naive-time method (#278).

* Removed the dependency on ellipsis in favor of the equivalent functions in
  rlang (#288).

* Updated tests related to writing UTF-8 on Windows and testthat 3.1.2 (#287).

* Updated all snapshot tests to use rlang 1.0.0 (#285).

* tzdb >=0.3.0 is now required to get access to the latest time zone database
  information (2022a).

* vctrs >=0.4.1 and rlang >=1.0.4 are now required (#297).

* cpp11 >=0.4.2 is now required to ensure that a fix related to unwind
  protection is included.

* R >=3.4.0 is now required. This is consistent with the standards of the
  tidyverse.

# clock 0.6.0

* New `date_count_between()`, `calendar_count_between()`, and
  `time_point_count_between()` for computing the number of units of time between
  two dates (i.e. the number of years, months, days, or seconds). This has a
  number of uses, like computing the age of an individual in years, or
  determining the number of weeks that have passed since the start of the year
  (#266).

* Modulus is now defined between a duration vector and an integer vector
  through `<duration> %% <integer>`. This returns a duration vector containing
  the remainder of the division (#273).

* Integer division is now defined for two duration objects through
  `<duration> %/% <duration>`. This always returns an integer vector, so be
  aware that using very precise duration objects (like nanoseconds) can easily
  generate a division result that is outside the range of an integer. In that
  case, an `NA` is returned with a warning.

# clock 0.5.0

* New `date_time_parse_RFC_3339()` and `sys_time_parse_RFC_3339()` for parsing
  date-time strings in the
  [RFC 3339](https://datatracker.ietf.org/doc/html/rfc3339) format. This format
  is a subset of ISO 8601 representing the most common date-time formats seen in
  internet protocols, and is particularly useful for parsing date-time strings
  returned by an API. The default format parses strings like
  `"2019-01-01T01:02:03Z"` but can be adjusted to parse a numeric offset from
  UTC with the `offset` argument, which can parse strings like
  `"2019-01-01T01:02:03-04:30"` (#254).

* To align more with RFC 3339 and ISO 8601 standards, the default formats used
  in many of the date formatting and parsing functions have been slightly
  altered. The following changes have been made:

  * Date-times (POSIXct):

    * `date_format()` now prints a `T` between the date and time.

    * `date_time_parse_complete()` now expects a `T` between the date and time
      by default.

  * Sys-times:

    * `format()` and `as.character()` now print a `T` between the date and time.

    * `sys_time_parse()` now expects a `T` between the date and time by default.

  * Naive-times:

    * `format()` and `as.character()` now print a `T` between the date and time.

    * `naive_time_parse()` now expects a `T` between the date and time by
      default.

  * Zoned-times:

    * `format()` and `as.character()` now print a `T` between the date and time.

    * `zoned_time_parse_complete()` now expects a `T` between the date and time
      by default.

  * Calendars:

    * `format()` and `as.character()` now print a `T` between the date and time.

    * `year_month_day_parse()` now expects a `T` between the date and time by
      default.

* Further improved documentation of undefined behavior resulting from attempting
  to parse sub-daily components of a string that is intended to be parsed into
  a Date (#258).

* Bumped required minimum version of tzdb to 0.2.0 to get access to the latest
  time zone database information (2021e) and to fix a Unicode bug on Windows.

# clock 0.4.1

* Updated a test related to upcoming changes in testthat.

# clock 0.4.0

* New `date_start()` and `date_end()` for computing the date at the start or
  end of a particular `precision`, such as the "end of the month" or
  the "start of the year". These are powered by `calendar_start()` and
  `calendar_end()`, which allow for even more flexible calendar-specific
  boundary generation, such as the "last moment in the fiscal quarter" (#232).

* New `invalid_remove()` for removing invalid dates. This is just a wrapper
  around `x[!invalid_detect(x)]`, but works nicely with the pipe (#229).

* All clock types now support `is.nan()`, `is.finite()`, and `is.infinite()`.
  Additionally, duration types now support `abs()` and `sign()` (#235).

* tzdb 0.1.2 is now required, which fixes compilation issues on RHEL7/Centos
  (#234).

# clock 0.3.1

* Parsing into a date-time type that is coarser than the original string is now
  considered ambiguous and undefined behavior. For example, parsing a string
  with fractional seconds using `date_time_parse(x)` or
  `naive_time_parse(x, precision = "second")` is no longer considered correct.
  Instead, if you only require second precision from such a string, parse the
  full string, with fractional seconds, into a clock type that can handle them,
  then round to seconds using whatever rounding convention is required for your
  use case, such as `time_point_floor()` (#230).

  For example:

  ```
  x <- c("2019-01-01 00:00:59.123", "2019-01-01 00:00:59.556")

  x <- naive_time_parse(x, precision = "millisecond")
  x
  #> <time_point<naive><millisecond>[2]>
  #> [1] "2019-01-01 00:00:59.123" "2019-01-01 00:00:59.556"

  x <- time_point_round(x, "second")
  x
  #> <time_point<naive><second>[2]>
  #> [1] "2019-01-01 00:00:59" "2019-01-01 00:01:00"

  as_date_time(x, "America/New_York")
  #> [1] "2019-01-01 00:00:59 EST" "2019-01-01 00:01:00 EST"
  ```

* Preemptively updated tests related to upcoming changes in testthat (#236).

# clock 0.3.0

* New `date_seq()` for generating date and date-time sequences (#218).

* clock now uses the tzdb package to access the date library's API. This
  means that the experimental API that was to be used for vroom has been
  removed in favor of using the one exposed in tzdb.

* `zone_database_names()` and `zone_database_version()` have been removed in
  favor of re-exporting `tzdb_names()` and `tzdb_version()` from the tzdb
  package.

# clock 0.2.0

* clock now interprets R's Date class as _naive-time_ rather than _sys-time_.
  This means that it no longer assumes that Date has an implied time zone of
  UTC (#203). This generally aligns better with how users think Date should
  work. This resulted in the following changes:

  * `date_zone()` now errors with Date input, as naive-times do not have a
    specified time zone.

  * `date_parse()` now parses into a naive-time, rather than a sys-time, before
    converting to Date. This means that `%z` and `%Z` are now completely
    ignored.

  * The Date method for `date_format()` now uses the naive-time `format()`
    method rather than the zoned-time one. This means that `%z` and `%Z` are
    no longer valid format commands.

  * The zoned-time method for `as.Date()` now converts to Date through an
    intermediate naive-time, rather than a sys-time. This means that the
    printed date will always be retained, which is generally what is expected.

  * The Date method for `as_zoned_time()` now converts to zoned-time through
    an intermediate naive-time, rather than a sys-time. This means that the
    printed date will always attempt to be retained, if possible, which is
    generally what is expected. In the rare case that daylight saving time makes
    a direct conversion impossible, `nonexistent` and `ambiguous` can be used
    to resolve any issues.

* New `as_date()` and `as_date_time()` for converting to Date and POSIXct
  respectively. Unlike `as.Date()` and `as.POSIXct()`, these functions always
  treat Date as a naive-time type, which results in more consistent and
  intuitive conversions. Note that `as_date()` does conflict with
  `lubridate::as_date()`, and the lubridate version handles Dates differently
  (#209).

* Added two new convenient helpers (#197):

  * `date_today()` for getting the current date (Date)

  * `date_now()` for getting the current date-time (POSIXct)

* Fixed a bug where converting from a time point to a Date or POSIXct could
  round incorrectly (#205).

* Errors resulting from invalid dates or nonexistent/ambiguous times are now
  a little nicer to read through the usage of an info bullet (#200).

* Formatting a naive-time with `%Z` or `%z` now warns that there were
  format failures (#204).

* Fixed a Solaris ambiguous behavior issue from calling `pow(int, int)`.

* Linking against cpp11 0.2.7 is now required to fix a rare memory leak issue.

* Exposed an extremely experimental and limited C++ API for vroom (#322).

# clock 0.1.0

* Added a `NEWS.md` file to track changes to the package.