FYI 2.5.0-rc-2 breaks Gant. Specifically, it's caused by changing groovy.util.CliBuilder to use Picocli
at sun.reflect.NativeMethodAccessorImpl.invoke0(Native Method)
Caused by: groovy.lang.ReadOnlyPropertyException: Cannot set readonly property: parser for class: groovy.util.CliBuilder
at gant.Gant$processArgs.call(Unknown Source)
... 6 more
The line in Gant is
Was this breakage intentional? I think a lot of stuff will break with parser not being able to be set anymore.
In 2.5.0-rc-2, groovy.util.CliBuilder delegates to groovy.cli.picocli.CliBuilder. The error is that the `parser` property of this class is no longer writable.
You can resolve this with 2.5.0-rc-2 by either not setting the `parser` property in the CliBuilder constructor or using the groovy.cli.commons.CliBuilder instead.
On the Groovy side I’m not sure what the best way is to make the transition easier. The picocli version of CliBuilder can not make use of the Commons-CLI parser class. We could modify CliBuilder to silently ignore the specified parser. (We’d have to rename the picocli ParserSpec `parser` property in CliBuilder to something else.)
I think I found a way to fix this.
This PR adds CliBuilder.setParser and CliBuilder.setFormatter methods that ignore the specified value and print a warning to stderr.
On Sun, May 6, 2018 at 9:14 PM, Remko Popma <[hidden email]> wrote:
I'll take a look. I'll also see if there are other places that can break.
If there are several ways it can break, then maybe we should not use a delegate to picocli and instead have folks switch the CliBuilder instance they instantiate.
Both CliBuilder implementations are there via their full package names, so encouraging people to use the non-delegated implementation of their choice is a good option. They will need to get used to such changes in Groovy 3 anyway.
For Groovy 3, we might have a way for the compiler to translate from the existing package names to the new ones brought about by JDK9 modules, in which case we'd do that for CliBuilder and we'd want to point it to the picocli package name by then (though if we do the change below we perhaps wouldn't include CliBuilder in the translations list).
For 2.5, we don't have the package translation in place and don't intend to. Because of that, the point of the delegation version is to aid in migration.Maybe it was too big a jump to try to delegate to the Picocli delegation in 2.5. If we find too many problems I think I should revert that change and delegate to the commons cli implementation instead but mark the whole delegation version class as deprecated.
Let me know what people think.
On Mon, May 7, 2018 at 3:45 PM, Keegan Witt <[hidden email]> wrote:
I realize now that simply removing the `parser` and `formatter` properties was a mistake. While the picocli version of CliBuilder cannot use commons-cli classes I should have anticipated that some applications are setting these properties.
I believe that I have a solution that provides a smoother migration path. PR
This should significantly reduce the disruption for applications like Gant that depend on the ability to modify these commons-cli properties in CliBuilder.
Would it be possible to have another 2.5-rc release with the above change and see what the community feedback is before we decide to revert the CliBuilder delegation to point to the commons-cli implementation?
Marking the delegation version class as deprecated is probably a good idea regardless of where it points to.
Short answer is we can have as many RCs as needed to get it right but we don't really want to be chopping and changing too much. Big changes are supposed to be settling down by now - fate just brought this change late in the game compared to what we'd like. I'd really like to get +ve feedback from a SNAPSHOT with that change before doing another release including it. I can merge in a couple of days time after conference commitments.
On Tue, May 8, 2018 at 2:27 PM, Remko Popma <[hidden email]> wrote:
Here's all the differences between 2.4 and 2.5
I'd guess setFormatter() wouldn't be a very common thing to do (I think it'd only really come up if HelpFormatter were subclassed for some reason), so that one's probably not a big risk. But I'm not sure setOptions() is safe to dummy up -- folks might be relying on this to configure the options to parse. This might be a reason not to default to picocli yet. What do you think?
I also looked at how current arguments passed on the commandline could break with the move, and the only thing I noticed was Commons CLI says it can do Java-like properties such as -Djava.awt.headless=true (although I didn't see how you'd actually configure these options unless you jammed all of them into 1 giant option).
On Tue, May 8, 2018 at 1:06 AM, Paul King <[hidden email]> wrote:
I updated the PR to make the signature of `setPosix` accept Boolean instead of boolean and added tests to ensure that nulls are handled correctly. Thanks for pointing that out!
The picocli version of CliBuilder has tests to verify that Java-like properties such as -Dkey=value are processed correctly in the new version just like in the commons-cli version. (For multiple properties the option can be specified multiple times: -Da=b -Dx=z )
I have not created a dummy setter for `setOptions` for the reason you mentioned:
if users use this to configure the options it is probably better to fail with a compiler error than silently ignoring such options.
It's hard to tell how many people would be impacted by this, but if we want 2.5 to be compatible with 2.4 we cannot make picocli the default in 2.5.
On Thu, May 10, 2018 at 8:42 AM, Keegan Witt <[hidden email]> wrote:
So do you want to change this back for RC3? I know it's late to change this, but arguably it was late to change it in RC2 too.
On Thu, May 10, 2018, 11:08 AM Remko Popma <[hidden email]> wrote:
|Free forum by Nabble||Edit this page|