On 16/04/2013, at 5:58 AM, Adam Murdoch <adam.murd...@gradleware.com> wrote:
> Hi,
>
> I think there's a deep flaw in this API: For almost every task there are
> different collections of inputs files that affect the output files and this
> API lumps them all together. Here are some examples:
>
> - The source files and the compile classpath of a Java compile task. I might
> implement this task so that it can deal with changes in the compile classpath
> but not the source files, so that it short-circuits compilation if the API of
> the compile classpath has not changed (regardless of how its packaged) but if
> it does compile anything, it recompiles everything. If a source file changes,
> it recompiles all the source.
>
> - The source files and the header files of a C++ compile task. I might
> implement this task so that it can deal with changes to source files but not
> header files. If a source file changes, I can recompile the output file for
> this source file. If a header file changes, I will recompile everything.
>
> - The source files and the compiler classpath of a Groovy or Scala compile
> task. When the compiler changes, I need to rebuild everything.
>
> - The source files, the classpath, and the config file of a check style task.
> For changes to source files, I can run the analysis for the changed files.
> For changes to the config file or the classpath, I need to reanalyse
> everything.
>
> - Unlikely but possible: A file moves from one collection to another
> collection, or is added to a second collection or removed from one collection
> but not all collections. In these cases, we're going to consider this file as
> unchanged, but it has changed role (it hasn't moved, it hasn't changed
> content, but its effect on the output is now different).
>
> - Additional input files declared outside the task via task.inputs.files().
> Changes to these should trigger a rebuild and should not be passed to the
> task.
>
> You get the idea. The API is going to have to allow a task to declare which
> collections of input files it can handle changes in, and which collections it
> cannot, and we're going to have to deliver those changes on a per collection
> basis. We may need this also for output files, but I don't see a use case for
> it yet.
No debate from me.
> How might this look? Some options:
>
> - We inject a change set object into the task action for each property you're
> interested in observing:
>
> @InputFile
> File configFile
>
> @InputFiles
> Set<File> classpath
>
> @TaskAction
> void doStuff(@ChangesFor('configFile') InputFileChanges configFileChanges,
> @ChangesFor('classpath') InputFilesChanges classpathChanges) { …
> }
>
This doesn't scream elegance. I think we should avoid this if we can.
> - We invoke an optional method for each property that you're interested in
> observing before we invoke the task action:
>
> // accepts the changes for the configFile property
> void configFileChanges(InputFileChanges changes) { … }
>
> // accepts the changes for the classpath property
> void classpathChanges(InputFilesChanges changes) { … }
>
> @TaskAction
> void doStuff() {
> // use whatever state the above methods have kept
> }
>
> We'd probably invoke those methods only if there are changes in the
> respective property. The benefit of this approach is that the task has the
> opportunity to force a rebuild or skip execution based on certain types of
> changes to the value, before the work is started.
A downside is that unit testing becomes more difficult because you have to
simulate the order in which Gradle would call the methods. It's a strange style
of API, and requiring the user to track such state across invocations feels
like complexity.
>
> - You query some context passed into the task action:
>
> @TaskAction
> void doStuff(IncrementalTaskContext context) {
> def configFileChanges = context.getChangesFor('configFile')
> def classpathChanges = context.getChangesFor('classpath')
> // need to get all changes before doing anything with them
> ...
> }
This seems the most straightforward to me. I can craft the logic I need in a
way that I control as the implementor. It also seems the most evolvable.
It is of course less declarative and there is also no way to detect an ivar
name mismatch until runtime.
>
> - A variant of the previous is to split the task action into 2 methods: one
> which gathers the changes and makes a decision about what type of execution
> is required, and one which does the work (or two: one for incremental, one
> for rebuild, or one for cleanup and one to build).
>
> void beforeAction(IncrementalTaskContext context) {
> // decide what to do based on the changes and let Gradle know your
> decision
> }
>
> @TaskAction
> void doStuff() {
> // use whatever state the above methods have kept
> }
I have the same problem with this one with regards to tracking internal state.
Just too error prone.
> - You use some kind of observable type for those properties you want to know
> the history for:
>
> @InputFile
> ObservableInputFile configFile
>
> @InputFiles
> ObservableInputFiles classpath
>
> @TaskAction
> void doStuff() {
> // can ask configFile if it has changed since last time, etc
> }
>
> The DSL layer would take care of getting the values in and out of these types.
The last is my preference I think. Using ivar names as strings is too error
prone for me and this avoids it. To easy to rename something and not rename it
everywhere.
We might need a parallel hierarchy though (ObservableInputFiles,
ObservableInputFileTree etc.) which is a bit painful. However, it's a
reasonably bound set I guess. We'd have to experiment with this and see what
the real cost is in terms of understandability. That is, when a user sees this
type in the DSL ref, do they understand how to effectively set/get it.
It's also leaking task implementation detail. The user should rarely care that
these things are observable to the implementation, but with this approach it is
exposed. We could hide it in our DSL ref, but that feels a bit awkward.
If we could make the last one not inconvenience users of the task then I think
that's worth exploring. If not, I think the "You query some context passed into
the task action:" one is the way forward.
>
>
>
> On 28/03/2013, at 2:43 AM, Daz DeBoer <darrell.deb...@gradleware.com> wrote:
>
>> G'day
>>
>> Now in master is a pretty cool new feature: you can now implement an
>> 'incremental' task that is informed about exactly which input files
>> have changed when the task is out of date.
>> This is very useful for something like a C++ compile task, as it means
>> that only the changed files need to be recompiled, rather than the
>> entire set of inputs.
>>
>> I've got a 'draft' DSL functioning, and would appreciate any feedback
>> you guys have. Here's a sample:
>>
>> class IncrementalSync extends DefaultTask {
>> @InputFiles
>> def FileCollection src
>>
>> @OutputDirectory
>> def File destination
>>
>> @TaskAction
>> void execute(TaskInputChanges inputs) {
>> if (inputs.allOutOfDate) {
>> FileUtils.forceDelete(destination)
>> }
>>
>> inputs.outOfDate({
>> FileUtils.copyFile(change.file, targetFile(change.file))
>> } as Action)
>> .removed({
>> FileUtils.forceDelete(targetFile(change.file))
>> } as Action)
>> .process()
>> }
>>
>> def targetFile(def inputFile) {
>> new File(destination, change.file.name)
>> }
>> }
>>
>> Notes:
>> 1. The way to implement an incremental task is to add a
>> TaskInputChanges parameter to your @TaskAction method. This must be a
>> typed parameter, and currently TaskInputChanges is the only parameter
>> type we support (but there are plans to add more, like
>> TaskOutputChanges). The reason for using a typed parameter is that
>> this is the way the task tells us what it wants: I thought about an
>> annotated parameter, but it seems kind of pointless when the
>> annotation would imply the type anyway. (Perhaps we can add an
>> annotation-based marker at a later stage, if it helps).
>>
>> 2. There are 2 discrete ways we report incremental changes:
>> - If the _only_ change to the task execution state is changed input
>> files, then TaskInputChanges.allOutOfDate() will be false, and only
>> the added/changed/removed files will be notified to the
>> TaskInputChanges.outOfDate() and .removed() actions.
>> - In the case of non-file changes to task inputs (properties, task
>> class) and changes to task output files, then Gradle will consider all
>> input files to be out of date. In this case,
>> TaskInputChanges.allOutOfDate() will be true, and every input file
>> will be reported to the TaskInputChanges.outOfDate() action.
>>
>> 4. The reason for the chained action methods combined with a final
>> process() method is that this allows us to stream changed inputs in
>> any order, and does not require us to persist these changes for a
>> subsequent method call. This is a little awkward, but doesn't force us
>> to jump through hoops. We could implement a more discrete API on top,
>> but it may be less efficient.
>>
>> 5. I haven't yet got any DSL magic applied to the TaskInputChanges
>> instance, so using a closure directly isn't (yet) possible. Not sure
>> how important that is for this DSL, or how tricky it will be to add.
>>
>> You can read more about the plans here:
>> https://github.com/gradle/gradle/blob/master/design-docs/incremental-build.md
>> Next steps for incremental tasks include providing access to changed
>> outputs and properties (in the case a task can handle these more
>> efficiently), automatically cleaning up stale outputs, and fixing some
>> bugs around the incremental nature of Copy tasks (and others).
>>
>> --
>> Darrell (Daz) DeBoer
>> Principal Engineer, Gradleware
>> http://www.gradleware.com
>> Join us at the Gradle Summit 2013, June 13th and 14th in Santa Clara,
>> CA: http://www.gradlesummit.com
>>
>> ---------------------------------------------------------------------
>> To unsubscribe from this list, please visit:
>>
>> http://xircles.codehaus.org/manage_email
>>
>>
>
>
> --
> Adam Murdoch
> Gradle Co-founder
> http://www.gradle.org
> VP of Engineering, Gradleware Inc. - Gradle Training, Support, Consulting
> http://www.gradleware.com
>
> Join us at the Gradle Summit 2013, June 13th and 14th in Santa Clara, CA:
> http://www.gradlesummit.com
>
--
Luke Daley
Principal Engineer, Gradleware
http://gradleware.com
Join me at the Gradle Summit 2013, June 13th and 14th in Santa Clara, CA:
http://www.gradlesummit.com
---------------------------------------------------------------------
To unsubscribe from this list, please visit:
http://xircles.codehaus.org/manage_email
