Parallel resolution of artifacts

I like the way this is going, but if I am not mistaken it appears that the patch ignores certain settings such as -Dmaven.repo.local and mirror settings in ~/.m2/settings.xml.

How so? The patch simply schedules the artifact resolution differently, but the same resolution code should work.

Updated patch that fixes a few concurrency issues, adds support for bigger builds, updates all wagon deps to rc1-SNAPSHOT

Updated patch (version 3) that works with Java 1.4 and eliminates problem with commons-logging (depends on WAGONHTTP-17)

Hey Don, you are right, it isn't your patch that broke this. I was testing this patch against 2.0.9-SNAPSHOT from SVN and got your changes intermingled with some recent updates from SVN which broke maven.repo.local setting.

Hey Chris - have you created a new issue for the problems you are experiencing?

Don, can you clarify the 3rd patch? It seems that you are using the normal wagon-http, not the standalone one from WAGONHTTP-17 - is this because you just substituted it in your local repo, or did you actually get it working with the standard dependency?

I substituted it in my local repo. In my latest -db release, I changed the version to be more clear.

do you have a test project that fails if you use standard wagon-http, but works with your version? The ones I've tested worked without making the substitution and I want to check the shade plugin is working as expected... you mentioned the apt plugin?

Try struts2-core - https://svn.apache.org/repos/asf/struts/struts2/trunk/core It uses the APT plugin, which depends on commons-logging. Another one I've heard of is the xdoclet plugin.

that worked. I'll be integrating this on another branch since some modifications are still needed and I noticed your own branch has moved on already

branch is at https://svn.apache.org/repos/asf/maven/sandbox/branches/maven/MNG-3379

Here are the changes I'd like to review:

• formatting
• are RuntimeExceptions appropriate? Seems more like they should be normal artifact resolution failures and the list of missing artifacts stuck together
• may need a different progress monitor since this shows no progress any more (one that can be done in parallel also is great)
• I'd like to investigate controlling this by settings (pool size)
• needs to be able to be merged to Maven trunk somehow (and check compatibility with embedder)

• are RuntimeExceptions appropriate? Seems more like they should be normal artifact resolution failures and the list of missing artifacts stuck together
  Agreed.

• may need a different progress monitor since this shows no progress any more (one that can be done in parallel also is great)
  Yeah, this is the last biggish chunk of work still to do. The progress bar just goes crazy

• I'd like to investigate controlling this by settings (pool size)
  Makes sense, but in my testing, it didn't see to help much after 5, but that could just be due to my particular multi-core box.

on the last one - it was more to be able to limit the number if you don't want to soak your bandwidth too

Actually, Http Client will already limit your connections to two per server, so you shouldn't have to worry. That might be a good one to make configurable as well.

might want to adjust the shading mechanism and do that in the distribution rather than the wagonhttp-17 method

do we need to be concerned at all with file-locking in the repository with this? We have an issue open for concurrent writes into the local repo, MNG-2802, but I expect this to be more about concurrent Maven processes running on the same machine, rather than concurrent retrieval of artifacts within one process. This is mainly because the resolution process will probably not cause the type of file overlap that could cause problems, though in the case of metadata I can see how it might...

No, you don't need to be worried about file-locking. The retrieval algorithm only retrieves artifacts from different groups in parallel, but ones from the same group in serial, which is where you would need file-locking.

I recreated the MNG-3379 branch but without the wagon changes. I also improved the console monitor so it now properly handles multiple files being retrieved simultaneously. What more needs to be done before the changes are moved into the 2.0.x branch?

the branch for this issue looks ok to me - though I see you've made additional changes on a labeled branch as well.

Is the MNG-3379 branch the complete list of changes you want to submit?

Yes, the changes on the other branch mostly have to do with backporting the fix to the 2.0.9 tag due to Resource not having a proper hashCode implementation back then.

got it - thanks

This issue passed the core ITs for 2.0.9 (and most for 2.0.10), i would like to see some more ITs testing the parallel downloads specifically and then we could consider it for .11

Brian - what kind of ITs do you envisage for this? Something that downloads a large number of artifacts from a file:// repository and check they all arrive safely?

Brett, yes something that would exercise this code. I've been running this branch locally for a few days and haven't noticed any issues yet. Bumping it back for consideration in 2.0.11

Hi Don,

(reposting my email to you here, since I never got a reply)

I was wondering if you could help us get this feature ready for introduction into Maven 2.1.0. We're putting together a release plan for it right now at:

http://docs.codehaus.org/display/MAVEN/Maven+2.1.0+Release+Plan

As you can see, things have changed WRT the versioning in Maven's future. The former 2.1 code - on trunk - has been renamed to 3.0-SNAPSHOT since it contains extensive changes, and we've decided to introduce some intermediary releases that will incorporate less risky new features such as parallel downloads that don't fit on the 2.0.x codeline appropriately. We're also going to be putting 2.0.x in end-of-life mode very soon.

I've heard good things about your patch for parallel downloads, but since I'm trying to push to make sure we have adequate documentation and tests for everything new that goes into Maven, we're still coming up short on this feature.

Can you provide us with a small design document just laying out the problem you solved, including scoping (for instance discussion about groupId limitations, and maybe any known issues on CI systems where multiple Maven processes may be sharing a single local repository with large numbers of parallel downloads happening simultaneously) and a short blurb describing your solution.

Also, we need to create some tests that we can use to guard against regressions in the future. Since you know this code the best, it'd be very helpful if you could help us design some test cases that might verify key pieces of the code to make sure nothing is broken (for instance, nothing that might introduce threading problems or file-locking issues).

If you can provide some of this information, it will allow us to put this feature into 2.1.0. Otherwise, I'm not sure whether we have the resource-hours available to grok the code, reconstruct the design decisions you took, and come up with adequate tests.

Starting consolidation of all 2.1.0-M* versions into 2.1.0-M1, which I'll then rename into 2.1.0

When we put this in, let's make sure we've documented the opt-out setting prominently for users.

even better, configure number of parallel threads, just set it to 1 to turn it off?

+1, sounds like a natural config to me.

initial commit done, now adding configuration

Way to go, Brett!

We need to follow up on the ConcurrentModificationException that's happening intermittently under stress. The discussion is here:

http://www.nabble.com/Synchronization-issue-with-parallel-downloads-td22136382.html

I was able to reproduce this once, I'm looking into ways to test it more reliably, and also checking all the code that gets touched for sync issues.

reviewing coverage details for artifact resolution, I've narrowed it down to just DefaultWagonManager that is not thread safe (but is expected to be)

the only potential problem I could find under current usage is the one triggered here, which is a bug in Plexus but can be externally synchronized. This wouldn't have tripped Don up as his branch was taken before the additional component configuration step was taken (for setting the HTTP headers).

I'm working on a test case in Maven for this and will then synchronize it properly.

Other potential problem spots: a number of configuration maps in DefaultWagonManager that are only modified on initialization (if they were modified during a multithreaded scenario there'd be a problem). Also, the repository list passed in to the resolution must not be modifiable as it isn't cloned.

sync issue fixed. Added a test case that can reproduce in most scenarios though it does rely on timing as Plexus is not testable in that fashion at the moment

Benjamin has identified an issue where exceptions in the resolution can cause the build to hang instead of fail correctly.

I'll work on a test case and fix for this tomorrow.