[
https://issues.apache.org/jira/browse/HADOOP-13223?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15319513#comment-15319513
]
Chris Nauroth commented on HADOOP-13223:
----------------------------------------
bq. It's not clear to me why a DLL would be less prone to path problems than an
EXE. It seems like we should just be putting a version number on the EXE, so
that we avoid these conflicts. We have the same problem with libhadoop-- see
HADOOP-11127.
You're correct that hadoop.dll suffers the same challenges as libhadoop.so, but
on Windows, the challenges are even greater.
First, there is the simple matter that there are 2 binaries to grapple with
instead of 1. Both hadoop.dll and winutils.exe are required.
Second, there is the problem of understanding how a Hadoop process loads
winutils.exe. In the case of hadoop.dll, the running process uses well-known,
well-defined dynamic linking mechanisms to find the dll. Experienced Windows
developers and admins will be familiar with the [DLL search
path|https://msdn.microsoft.com/en-us/library/windows/desktop/ms682586(v=vs.85).aspx].
For winutils.exe, there is no such familiar ground on which developers and
admins can build an understanding. Hadoop uses separate, arbitrary logic to
find the winutils.exe binary. Unlike the DLL search path, this is not
consistent with typical dynamic linking practices, so it can be a source of
confusion.
I am +1 for migrating more functionality into hadoop.dll and eventually
eliminating winutils.exe. This addresses the additional difficulty of
coordinating 2 binaries and the additional difficulty of understanding how it
gets loaded. It does not address the challenge of version compatibility
between Java and native code during dynamic linking, but that issue is tracked
elsewhere.
> winutils.exe is a bug nexus and should be killed with an axe.
> -------------------------------------------------------------
>
> Key: HADOOP-13223
> URL: https://issues.apache.org/jira/browse/HADOOP-13223
> Project: Hadoop Common
> Issue Type: Improvement
> Components: bin
> Affects Versions: 2.6.0
> Environment: Microsoft Windows, all versions
> Reporter: john lilley
>
> winutils.exe was apparently created as a stopgap measure to allow Hadoop to
> "work" on Windows platforms, because the NativeIO libraries aren't
> implemented there (edit: even NativeIO probably doesn't cover the operations
> that winutils.exe is used for). Rather than building a DLL that makes native
> OS calls, the creators of winutils.exe must have decided that it would be
> more expedient to create an EXE to carry out file system operations in a
> linux-like fashion. Unfortunately, like many stopgap measures in software,
> this one has persisted well beyond its expected lifetime and usefulness. My
> team creates software that runs on Windows and Linux, and winutils.exe is
> probably responsible for 20% of all issues we encounter, both during
> development and in the field.
> Problem #1 with winutils.exe is that it is simply missing from many popular
> distros and/or the client-side software installation for said distros, when
> supplied, fails to install winutils.exe. Thus, as software developers, we
> are forced to pick one version and distribute and install it with our
> software.
> Which leads to problem #2: winutils.exe are not always compatible. In
> particular, MapR MUST have its winutils.exe in the system path, but doing so
> breaks the Hadoop distro for every other Hadoop vendor. This makes creating
> and maintaining test environments that work with all of the Hadoop distros we
> want to test unnecessarily tedious and error-prone.
> Problem #3 is that the mechanism by which you inform the Hadoop client
> software where to find winutils.exe is poorly documented and fragile. First,
> it can be in the PATH. If it is in the PATH, that is where it is found.
> However, the documentation, such as it is, makes no mention of this, and
> instead says that you should set the HADOOP_HOME environment variable, which
> does NOT override the winutils.exe found in your system PATH.
> Which leads to problem #4: There is no logging that says where winutils.exe
> was actually found and loaded. Because of this, fixing problems of finding
> the wrong winutils.exe are extremely difficult.
> Problem #5 is that most of the time, such as when accessing straight up HDFS
> and YARN, one does not *need* winutils.exe. But if it is missing, the log
> messages complain about its absence. When we are trying to diagnose an
> obscure issue in Hadoop (of which there are many), the presence of this red
> herring leads to all sorts of time wasted until someone on the team points
> out that winutils.exe is not the problem, at least not this time.
> Problem #6 is that errors and stack traces from issues involving winutils.exe
> are not helpful. The Java stack trace ends at the ProcessBuilder call. Only
> through bitter experience is one able to connect the dots from
> "ProcessBuilder is the last thing on the stack" to "something is wrong with
> winutils.exe".
> Note that none of these involve running Hadoop on Windows. They are only
> encountered when using Hadoop client libraries to access a cluster from
> Windows.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
