Changes for version 0.315 - 2026-09-25
- Windows
- On MSWin32 the module runs its script with the first of `python`, `py -3` and `python3` that reports itself as Python 3. It ran `python3`, which a standard Windows install from python.org does not have, so on Strawberry Perl every call that ran its script failed wherever Python *was* installed. It never reached a CPAN Testers report, because the Windows smokers have no Python and the tests that need it are skipped there. The version is checked rather than the name trusted, so the Microsoft Store stub that Windows 10 and 11 install as `python.exe` is passed over. Everywhere else the command is still `python3`. With no interpreter, `plt` dies naming the ones it tried; `execute => 0` still writes the script.
- A script that fails is reported under the command that actually ran it, not always as `python3`.
- The test suite asks the module which Python to use instead of naming `python3`, so a Windows smoker that does have Python now tests the module with it rather than skipping. The parser checks no longer pass double quotes on the command line, which Perl on Win32 does not escape. `t/08.python.interpreter.t` drives the Windows search on unix, with stand-in interpreters on a private `PATH`.
- Figures that were drawn wrong
- `violin` whiskers are computed from sorted data. The whisker code, taken from matplotlib's own violin example, reads the smallest and largest values as the first and last of each group; matplotlib's example sorts its data first and this module did not, so any group not already in ascending order got the wrong whiskers. For `[10, 9, ..., 1, 100]` the lower whisker was clipped to the first quartile and not drawn at all.
- `violin` shapes are smooth again for small groups. `violinplot`'s `points` -- how many points the density curve is evaluated at -- was set to the size of the smallest group, so a group of 5 values drew every violin in the figure as a 5-sided polygon. It is now left at matplotlib's default of 100.
- `colored_table` with `cb_logscale` colors its cells on a log scale. Only the colorbar was logarithmic: the cells were colored through a linear scale whatever the option said, so a table of 1, 10, 100 and 1000 drew three of its four cells in nearly the same color beside a colorbar that said otherwise. Cells and colorbar now share one scale. A log scale starts at the smallest cell above 0, as it did when matplotlib chose it, and a `cb_min` of 0 or below is refused, since a log scale cannot start there.
- `hist` takes any matplotlib color. Only a plain name such as `red` was quoted, so `color => { A => '#ff0000' }` wrote `color = #ff0000` and python read the rest of the line as a comment -- a syntax error -- while `tab:blue` and a grayscale `'0.5'` went in as python rather than as colors.
- `hist` draws a single `color` for every set, as documented. Only the hash of one color per set was ever read, so `color => 'red'` was accepted and the bars stayed blue. An array of colors is refused with a pointer to the hash form, since `hist` has no order of sets for an array to follow.
- A backslash in a title or label reaches matplotlib as written. Text passed as `title`, `xlabel`, `ylabel`, `suptitle` and the rest was put into a python string without escaping, so the mathtext `'$\alpha$'` arrived with a bell character in it and `'C:\new'` with a newline. Text that already carries quotes of its own is still passed through untouched.
- The `suptitle` of a figure of several subplots is quoted like a single plot's. `suptitle => 'Two groups: mean.'` wrote `plt.suptitle(Two groups: mean.)`, a syntax error, and non-ASCII text given as raw bytes was not decoded.
- `wide` draws its groups in the sorted order of their names, so the legend no longer changes order from one run to the next, and groups with no `color` of their own take successive colors of matplotlib's default cycle (`C0`, `C1`, ...). They were all `b` (pure blue), so the groups could not be told apart without a `color` hash. This changes the picture for anyone relying on that default.
- plt no longer changes what it is given
- Calling `plt({ ... })` twice with the same hash draws the same figure twice. `plt` wrote into the hash it was given: `add` was deleted once drawn, `title` renamed `set_title`, and subplots told whether to draw a colorbar, so the second figure silently lost its overlays. The wrapper functions (`bar`, `plot`, ...) already copied the top level; now every call copies the top level, each subplot and each `add` graph.
- `imshow` no longer replaces the strings in the caller's own rows with integers, and a string that `stringmap` does not name is refused rather than drawn as a blank cell.
- `bar`/`barh` no longer write 0 into the holes of the caller's grouped data, and no longer overwrite a `label` array given with a hash of hashes. A `p` subplot given as an array no longer appends its overlays to the base plot's own `add` array.
- Data that cannot be drawn is refused, naming what is wrong
- `plot` with hash data refuses a value that is not a number, as the array form always has. `data => { A => [[1,2],['a',3]] }` wrote `y = [a,3]` and died as a `NameError` in the temporary file.
- A `colored_table` cell that is not a number (`'NA'`, say) is refused, naming the cell, instead of being written as `d.append([1,NA])`. A cell with no value is still drawn in `undef.color`.
- `scatter` refuses non-numeric values in any axis, including the color axis, which was not checked at all; only undefined x and y values were refused before.
- `hexbin` checks its data, which it never had: an undefined or non-numeric point dies naming the key and the indices. A non-numeric `vmin`, `vmax` or `mincnt` is refused.
- Bug fixes
- A multi-set `scatter` may mix sets with and without a color axis. The color key of one set was kept for every set after it, so a set of only `x` and `y` following one with `z` died as `"z" isn't defined for set "B"`.
- `hexbin` with `cb_logscale` accepts `vmin` and `vmax`, which now go to the log scale itself. Passed beside it they made matplotlib raise `Passing a Normalize instance simultaneously with vmin/vmax is not supported`; `hist2d` has handled this since at least 0.3.
- The diagnostics printed before `hexbin`, `hist2d` and multi-set `scatter` die, and the usage examples printed when `plt` is called without enough to draw, go to STDERR instead of STDOUT, with every other diagnostic.
- Testing
- `t/07.review.fixes.t` covers each fix above: the generated python is read as text everywhere, parsed where python3 is present, and where matplotlib is present the violin whiskers, the colored_table norm and the hexbin bounds are read back from the running script. `Storable`, which it uses to check that the caller's data is left alone, is declared as a test prerequisite; it is core.
Modules
Access Matplotlib from Perl; providing consistent user interface between different plot types