Changes for version 0.3423 - 2026-11-15
- Distribution
- lib/Stats/LikeR.pm.rej, a reject left by patch(1) after its hunk had been applied by hand, is no longer in the tarball. It was tracked in git, [PruneFiles] keeps all of lib/, and so 0.3422 shipped it, and `make install` copied it into site_perl beside the module. The module itself was unaffected. .gitignore now ignores *.rej, and dzil.sh leaves *.rej out of what it adds and refuses a tarball that has one, as it has done for *.orig since 0.3415 and 0.3416 shipped lib/Stats/LikeR.pm.orig.
- read_table()
- read_table() with keep_cols or drop_cols no longer aborts perl on a perl built with -DDEBUGGING. The parser stands a NULL in its row buffer for each field whose column is left out, and when the buffer was full it stored that NULL with av_push(), which such a perl asserts against ("Perl_av_push: Assertion `val' failed"). It now grows the buffer with av_extend() and stores the NULL itself, as it already did while there was room. A perl without -DDEBUGGING stored the NULL correctly and was never affected. The module was at fault, not the tests: t/read_table.keep_cols.t died before its first subtest and t/read_table.drop_cols.t after its 17th or 33rd, in all 11 FAIL reports for 0.3422, every one from a -DDEBUGGING perl (5.30.1 to 5.45.4, double, long double and quadmath). A -DDEBUGGING perl 5.44.0 reproduced both failures here, at the same subtests, and passes the whole suite with the fix. That perl is now part of the test matrix this is released against.
- colClasses is no longer applied to a field the table never holds. When a header repeats a name, the last field of that name is the column, and a CSV read still converted the earlier ones as it cut them, so read_table() croaked over a value in a column it was about to throw away ('e,b,e' over 'x,2,3' with colClasses integer for both e). An .xlsx and the perl row closure already read that table as { e => 3, b => 2 }. An aoa keeps every field, so there the field is still converted and still refused.
- view()
- A hash key that perl holds downgraded, such as a UTF-8 "é" stored as the byte E9, is now shown as the UTF-8 string keys() returns. It was read from its stored bytes and printed as a lone E9 byte in a HoH's row labels and in the column names of an AoH, a HoA or a HoH, while a flat hash printed the same key as C3 A9.
- A gap that no buffer can hold, such as gap => 1e12, is now refused with "view: the table is too large to render" instead of ending perl with "Out of memory".
- Listing every key of a hash no longer makes one temporary copy per key that lived until view() returned.
- Data-frame helpers
- concat() and rbind() skip an empty frame of either kind, as their documentation says. An empty [] among HoAs, or {} among AoHs, was dereferenced as the other kind and died "Not a HASH reference".
- melt() keeps a name given twice in id_vars once. Each mention was pushed again, so a hoa's id column came out twice as long as the rest. A hoa is now built a column at a time, which melts 300,000 rows of two id and five value columns in 0.17 s instead of 0.62 s, and the other output types decide their shape once rather than once per cell.
- ffill() and bfill() with limit count only the cells they can write. A row that is no hash, or an AoA row too short for the column, was counted against the limit, so ffill([{a => 1}, undef, {a => undef}], limit => 1) left the last row unfilled.
- chunk() in scalar context gives 0 for empty input, not undef, and refuses a bad size or parts for empty input as for any other.
- aoh2hoh()'s errors for a row that is no hash and for a missing key now name the function, as its other errors do.
Modules
Get basic statistical functions, like in R, but with Perl using XS for performance