[jira] [Comment Edited] (HADOOP-11656) Classpath isolation for downstream clients

["Colin Patrick McCabe (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HADOOP-11656?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14349332#comment-14349332 ] 

Colin Patrick McCabe edited comment on HADOOP-11656 at 3/5/15 7:32 PM:
-----------------------------------------------------------------------

bq. Andrew wrote: One additional note related to this, we can spend a lot of time right now distributing 100s of MBs of jar dependencies when launching a YARN job. Maybe this is ameliorated by the new shared distributed cache, but I've heard this come up quite a bit as a complaint. If we could meaningfully slim down our client, it could lead to a nice win.

I'm frustrated that nobody responded to my earlier suggestion that we de-duplicate jars.  This would drastically reduce the size of our install, and without rearchitecting anything.

In fact I was so frustrated that I decided to write a program to do it myself and measure the delta.  Here it is:

Before:
{code}
du -h /h
249M    /h
{code}

After:
{code}
du -h /h
140M    /h
{code}

Seems like deduplicating jars would be a much better project than splitting into a client jar, if we really cared about this.

And here is the de-duplicator program I wrote (in Go):
{code}
package main

import (
  "path/filepath"
  "flag"
  "fmt"
  "os"
)

var basePathToFullPath map[string][]string = map[string][]string{ }

func visit(path string, f os.FileInfo, err error) error {
	if err != nil {
		panic(err)
	}
	if f.Mode().IsDir() {
		return nil
	}
	base := filepath.Base(path)
	bases := basePathToFullPath[base]
	if bases == nil {
		bases = make([]string, 0, 1)
	}
	bases = append(bases, path)
	basePathToFullPath[base] = bases
	fmt.Printf("%s -> %s\n", base, path)
	return nil
}

func main() {
	flag.Parse()
	if len(os.Args) < 2 {
		fmt.Printf("Usage: %s [path]\n", os.Args[0])
		os.Exit(1)
	}
	root := os.Args[1]
	err := filepath.Walk(root, visit)
	if err != nil {
		fmt.Printf("Error while traversing %s: %s\n", root, err.Error())
		os.Exit(1)
	}
	for basePath, fullPaths := range basePathToFullPath {
		if len(fullPaths) <= 1 {
			continue
		}
		absPath, err := filepath.Abs(fullPaths[0])
		if err != nil {
			fmt.Printf("failed to find abspath of %s: %s\n", fullPaths[0], err.Error())
			os.Exit(1)
		}
		fmt.Printf("Handling %s\n", basePath)
		for i := 1; i < len(fullPaths); i++ {
			fmt.Printf("rm %s\n", fullPaths[i])
			err = os.Remove(fullPaths[i])
			if err != nil {
				panic(err)
			}
			fmt.Printf("ln %s %s\n", absPath, fullPaths[i])
			err = os.Symlink(absPath, fullPaths[i])
			if err != nil {
				panic(err)
			}
		}
	}
}
{code}

The measurements I made were made against trunk (branch 3.0.0)


was (Author: cmccabe):
bq. Andrew wrote: One additional note related to this, we can spend a lot of time right now distributing 100s of MBs of jar dependencies when launching a YARN job. Maybe this is ameliorated by the new shared distributed cache, but I've heard this come up quite a bit as a complaint. If we could meaningfully slim down our client, it could lead to a nice win.

I'm frustrated that nobody responded to my earlier suggestion that we de-duplicate jars.  This would drastically reduce the size of our install, and without rearchitecting anything.

In fact I was so frustrated that I decided to write a program to do it myself and measure the delta.  Here it is:

Before:
{code}
du -h /h
249M    /h
{code}

After:
{code}
du -h /h
140M    /h
{code}

Seems like deduplicating jars would be a much better project than splitting into a client jar, if we really cared about this.

And here is the de-duplicator program I wrote:
{code}
package main

import (
  "path/filepath"
  "flag"
  "fmt"
  "os"
)

var basePathToFullPath map[string][]string = map[string][]string{ }

func visit(path string, f os.FileInfo, err error) error {
	if err != nil {
		panic(err)
	}
	if f.Mode().IsDir() {
		return nil
	}
	base := filepath.Base(path)
	bases := basePathToFullPath[base]
	if bases == nil {
		bases = make([]string, 0, 1)
	}
	bases = append(bases, path)
	basePathToFullPath[base] = bases
	fmt.Printf("%s -> %s\n", base, path)
	return nil
}

func main() {
	flag.Parse()
	if len(os.Args) < 2 {
		fmt.Printf("Usage: %s [path]\n", os.Args[0])
		os.Exit(1)
	}
	root := os.Args[1]
	err := filepath.Walk(root, visit)
	if err != nil {
		fmt.Printf("Error while traversing %s: %s\n", root, err.Error())
		os.Exit(1)
	}
	for basePath, fullPaths := range basePathToFullPath {
		if len(fullPaths) <= 1 {
			continue
		}
		absPath, err := filepath.Abs(fullPaths[0])
		if err != nil {
			fmt.Printf("failed to find abspath of %s: %s\n", fullPaths[0], err.Error())
			os.Exit(1)
		}
		fmt.Printf("Handling %s\n", basePath)
		for i := 1; i < len(fullPaths); i++ {
			fmt.Printf("rm %s\n", fullPaths[i])
			err = os.Remove(fullPaths[i])
			if err != nil {
				panic(err)
			}
			fmt.Printf("ln %s %s\n", absPath, fullPaths[i])
			err = os.Symlink(absPath, fullPaths[i])
			if err != nil {
				panic(err)
			}
		}
	}
}
{code}

The measurements I made were made against trunk (branch 3.0.0)

> Classpath isolation for downstream clients
> ------------------------------------------
>
>                 Key: HADOOP-11656
>                 URL: https://issues.apache.org/jira/browse/HADOOP-11656
>             Project: Hadoop Common
>          Issue Type: New Feature
>            Reporter: Sean Busbey
>            Assignee: Sean Busbey
>              Labels: classloading, classpath, dependencies, scripts, shell
>
> Currently, Hadoop exposes downstream clients to a variety of third party libraries. As our code base grows and matures we increase the set of libraries we rely on. At the same time, as our user base grows we increase the likelihood that some downstream project will run into a conflict while attempting to use a different version of some library we depend on. This has already happened with i.e. Guava several times for HBase, Accumulo, and Spark (and I'm sure others).
> While YARN-286 and MAPREDUCE-1700 provided an initial effort, they default to off and they don't do anything to help dependency conflicts on the driver side or for folks talking to HDFS directly. This should serve as an umbrella for changes needed to do things thoroughly on the next major version.
> We should ensure that downstream clients
> 1) can depend on a client artifact for each of HDFS, YARN, and MapReduce that doesn't pull in any third party dependencies
> 2) only see our public API classes (or as close to this as feasible) when executing user provided code, whether client side in a launcher/driver or on the cluster in a container or within MR.
> This provides us with a double benefit: users get less grief when they want to run substantially ahead or behind the versions we need and the project is freer to change our own dependency versions because they'll no longer be in our compatibility promises.
> Project specific task jiras to follow after I get some justifying use cases written in the comments.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)