Declare inputs and outputs to support up-to-date checking

[oehme]

The spotless check task currently doesn't have any inputs and outputs declared and is thus never considered up-to-date. This is slowing down user's builds unnecessarily.

[nedtwigg]

The input is the source code, and the output is also the exact same source code. Can gradle handle this case?

[oehme]

I'm only talking about the check task. The update task cannot be made incremental, for exactly the reason you said.

[nedtwigg]

My largest project has 6.87MB of java source, spread across 1636 files (in many subprojects, of course :). It takes 5.6 seconds to sum their filesizes with this code in my buildscript:

println 'numFiles = ' + files.size()
long start = System.currentTimeMillis()
long totalSize = 0;
for (File f : files) {
    totalSize += f.length()
}
long elapsed = System.currentTimeMillis() - start
println 'filesize_MB=' + totalSize / 1_000_000
println 'elapsed_ms=' + elapsed

Running spotlessCheck on these files with these rules:

spotless {
    java {
        licenseHeaderFile
        importOrderFile
        eclipseFormatFile
        custom 'AutoValue fix', { it.replace('autovalue.shaded.', '') }
    }
}

Takes 38 seconds with a fresh daemon, 16 seconds on the second run, and levels out around 14s.

Given that it takes 6 seconds just to sum their filesizes, I think adding incremental support can, at best, take ~8-10 seconds off of a very large check task, which is about half. It'll be much much faster than any reasonable set of unit tests on a codebase this size.

@oehme How does gradle determine up-to-date? Does it trust filesize+lastModified, some kind of inotify, or take a hash?

The hardest part will be serializing custom input functions. In the example above, if the user changes the AutoValue fix function, we want to be sure to recalculate the "clean" state. If we don't have a good way to do that, then we introduce a minor staleness bug. If spotlessCheck is very slow, and the up-to-date checking is very fast, then it would be worthwhile. But unless gradle's up-to-date checking is doing something very clever like hooking into filesystem events, it seems to me that it's not worth the complexity.

[oehme]

That code you showed is too simplistic. Gradle can up-to-date check projects with a 100000 source files in about 5s. We have performance tests for that. Also, in the future Gradle will even do this pre-emptively (through file system listeners), so that it already knows the up-to-date state before you even run the next build.

The problem is that the spotless check task will run even if the source code for a project did not change. Imagine you only changed one of your projects, but spotless still insists on re-running on every project. That's a big waste of time. There really is no good reason not to do up-to-date checking ;)

You can hash custom functions by hashing their byte code (which you can get from their ClassLoader).

[oehme]

I just saw that spotless adds a task to the root project which checks the complete source tree. This is so much unnecessary work on incremental builds. That calls for a redesign towards much more granular tasks. The plugin should add a task for every sourceSet. If the user wants to check stuff outside of the sourceSets, he can add another check task that only checks some specific files (let's say, the build.gradle files in his gradle folder).

Basically the extension defines how to check (including the file extension, but not the full path). The task decides what to check.

[nedtwigg]

spotless adds a task to the root project

It doesn't add a task to the root project, just the project. You can add it to the root project (I often do), or you can add it to each subproject independently if you prefer.

and then checks the complete source tree

It is true that, by default, it creates one spotlessJavaCheck task for all of a single project's sourceSets, rather than creating a task per sourceSet. That's because most people want to have the same format for all of their sourceSets. If the user wants separate formatting rules for different sourceSets, that's easy to do too, but I don't see a reason to make that the default behavior.

I don't understand why one task with 500 files would be faster than two tasks with 250 files each, with or without incremental build support, so I don't understand why this causes unnecessary work. But if it does, and we need to redesign, I'm open to that :)

The problem is that the spotless check task will run even if the source code for a project did not change. Imagine you only changed one of your projects, but spotless still insists on re-running on every project.

Imagine you want to calcualte the sum of 1000 numbers, and rather than caching each incremental sum, you just recalculate the whole sum when any number changes. Supporting incremental processing is only a win if the processing is more expensive than tracking the changes. Spotless is very simple, and so it is very fast - the bottleneck is disk IO.

Until we can measure that something is a bottleneck, and we can measure that we have sped it up, I don't think we can justify adding complexity. I'm open to the idea that my benchmark is naive, but I don't see how yet. My machine is an old laptop with a magnetic disk, and it takes 6 seconds to look at the metadata of 1600 files, and 14 seconds to read them from disk, format them, and check that they are clean. If your build server has an SSD that can read the metadata for 100,000 files in 5 seconds, maybe it's also powerful enough to format them in 10 seconds?

You can hash custom functions by hashing their byte code

That's a clever idea, I like it! When gradle implements filesystem listeners, I could maaaaybe see that it might be worth the complexity of implementing this feature. But for now, I don't understand why users would experience a faster build, because I don't understand how gradle can check for up-to-date in less time than my naive benchmark.

[oehme]

That's because most people want to have the same format for all of their sourceSets.

Sure, the format would be specified in the extension, which all tasks reuse. But each sourceSet should have its own task. When I only change my test code, I don't want to re-check the main code. It's a waste of time.

I don't understand why one task with 500 files would be faster than two tasks with 250 files each, with or without incremental build support

You are thinking way too much in terms of full builds. Gradle is all about incremental builds. When I change a single source file in a single project, especially with continuous build, I expect a response in a few hundred milliseconds, not several seconds or more.

Imagine you want to calcualte the sum of 1000 numbers, and rather than caching each incremental sum, you just recalculate the whole sum when any number changes. Supporting incremental processing is only a win if the processing is more expensive than tracking the changes. Spotless is very simple, and so it is very fast - the bottleneck is disk IO.

I won't get into an argument here, just test it yourself and you'll be suprised how much faster up-to-date checking is than you think

task foo() {
 inputs.files fileTree('someLargeDirectory')
 outputs.file file('dummy')
 doLast {
   //do nothing, just seeing how fast up-to-date checks are
 }
}

I did that on a 256MB tree containing 100000 source files. ./gradlew foo takes 1.3s.

Until we can measure that something is a bottleneck, and we can measure that we have sped it up, I don't think we can justify adding complexity.

As I showed you this is trivial to measure. Also, you are talking as if adding incremental build support was some monumental task. For a simple plugin like this, it's a weekend project at best.

[oehme]

Running spotless on that same filetree takes 13s when I change a single file. That's 10 times longer! Imagine how much faster it would be if it only checked that single file. Gradle makes that trivial.

The config was very simplistic, so this gets worse as you add more rules:

spotless {
    format 'native', {
        target '**/*.cpp', '**/*.c', '**/*.h'

        customReplace      'Not enough space after if', 'if(', 'if ('
        customReplaceRegex 'Too much space after if', 'if +\\(', 'if ('
    }
}

[nedtwigg]

Imagine how much faster it would be if it only checked that single file

We don't have to imagine, we can calculate it ;-) It would be ~12s seconds faster on a 256MB source tree. On a 26MB source tree it would presumably be ~1.2 seconds faster. The biggest source tree I use in my humble world is just 7MB, though my laptop seems to be a lot crummier than yours ;-)

If spotless is consuming 10% of a build's duration, then a 10x speedup in spotless is a 9% build speedup.

I'd be happy to merge a PR. To me, this data confirms that spotless isn't slow enough for this optimization to make it to the top of my weekend project priority queue, at least not on a "utility" basis. But it would be a fun project to learn about or demonstrate Gradle's incremental APIs! If anyone is interested in doing this project for that purpose, go for it!

After thinking about the custom function issue, it doesn't bother me that much anymore - it's fine if changing a custom function doesn't cause spotlessCheck to recalculate. Changes to these functions will almost certainly be tested using spotlessApply, so it's okay if spotlessCheck is broken in this narrow way.

[oehme]

I corrected the number above. It is 10s slower or in other words 10 times as slow.

And remember this is for a trivial set of replacements. I'm sure it's way worse if you add more. I'll give you some more numbers later.

[nedtwigg]

Currently, this is the only task in the project, FormatTask. If anybody has time and need to make the change, I'm happy to merge and release the PR. Clone this repo, run gradlew ide, and you've got a dev environment. Here's the gradle docs on incremental builds.

I'd guess that making this change might require creating a separate FormatApplyTask and FormatCheckTask task, but if there's a way to avoid that it would be better to be able to release as 2.1.0 rather than 3.0.0, but whatever the code needs :)

[oehme]

I will be happy to assist the contributor with this. It will definitely require a 3.0 to do this right.

We should fix some other non-idiomatic Gradle usages while doing this. For instance the formats DSL and how the plugin creates tasks from it is not very user friendly. There are lots of special characters needed like quotes, commas and parentheses. It just looks very different from other Gradle APIs. Also, the task is not available except using afterEvaluate which makes customization look awkward. This can easily be fixed by following the same patterns that all Gradle core plugins use for their DSL. And if we're breaking API to fix the tasks, we should take this into account too.

We should track that in its own issue, but assign it to the same 3.0 milestone. @nedtwigg can you set that up? :)

Here's my design for a Gradle-idiomatic DSL for spotless. The plugin would create a pair of check/apply tasks for every 'target'. This will make the plugin faster, more flexible and easier to extend.

spotless {
  //formats specify how to format certain kinds of code
  formats {
    // by using the named container API you get stuff like 'all' for free :)
    all {
     lineEndings 'GIT_ATTRIBUTES'
    }
    //java is a special method that returns a more specific type
    java {
      lineEndings 'UNIX'
    }
    //so is freshmark :)
    freshmark {

    }
   //custom formats look just like pre-defined ones, no Strings or commas
    misc {
      //the format only specifies the file extension it cares about
      //which filetrees to process is decided on the target level
      extensions '.gradle', '.md', '.gitignore'
      //use boolean assignment instead of no-arg methods to
      //avoid as many parentheses as possible and also making it
      //easy to turn off again
      trimTrailingWhitespace true
      indentWith '\t'
      endWithNewline true
      step('mySpecialStep') {
        //imperative code here
      }
    }
  }
  //targets select file trees to scan and optionally which formats to apply
  targets {
    buildFiles {
      dir(projectDir) //careful here, we don't want to recurse through the whole project ;)
      fileTree("${projectDir}/gradle")
      //optional, default is all formats
      formats formats.misc
    }
    //when the java plugin is applied, it automatically adds a target
    //for every sourceSet of the project
  }
}

[nedtwigg]

(Unrelated aside: setting the lineEndings property is almost always a bad idea now that spotless supports gitattributes, see here for explanation)

I've created issue #32 to discuss your proposed DSL change.

We can definitely implement up-to-date checking and incremental support without changing the DSL, they're almost unrelated. Even if we decide to break the existing DSL, we should first release a version which adds incremental support for people who can't immediately make the jump to 3.0.0.

Note to potential contributors: I will quickly merge and release any PR's which add up-to-date support or incremental build, but please don't break the DSL.

[oehme]

We will need to break up the tasks into two different types, so that will be a breaking change. But yes, we don't need to change the spotless extension DSL to get incremental builds.

Breaking up the DSL to create more fine grained tasks will make incremental builds even better though, that's why I mentioned it here.