[
https://issues.apache.org/jira/browse/SOLR-12290?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16465235#comment-16465235
]
Uwe Schindler commented on SOLR-12290:
--------------------------------------
This commit breaks TestCSVLoader on Windows. It looks like Solr no longer
closes any content streams that are passed separately to the servlet stream:
[junit4] 2> NOTE: reproduce with: ant test -Dtestcase=TestCSVLoader
-Dtests.method=testLiteral -Dtests.seed=B40869AE03F63CBC -Dtests.locale=ms
-Dtests.timezone=ECT -Dtests.asserts=true -Dtests.file.encoding=UTF-8
[junit4] ERROR 0.06s | TestCSVLoader.testLiteral <<<
[junit4] > Throwable #1: java.nio.file.FileSystemException: C:\Users\Uwe
Schindler\Projects\lucene\trunk-lusolr1\solr\build\solr-core\test\J0\temp\solr.handler.TestCSVLoader_B40869AE03F63CBC-001\TestCSVLoader-006\solr_tmp.csv:
Der Prozess kann nicht auf die Datei zugreifen, da sie von einem anderen
Prozess verwendet wird.
[junit4] > at
__randomizedtesting.SeedInfo.seed([B40869AE03F63CBC:B2ACAF677D87833E]:0)
[junit4] > at
sun.nio.fs.WindowsException.translateToIOException(WindowsException.java:86)
[junit4] > at
sun.nio.fs.WindowsException.rethrowAsIOException(WindowsException.java:97)
[junit4] > at
sun.nio.fs.WindowsException.rethrowAsIOException(WindowsException.java:102)
[junit4] > at
sun.nio.fs.WindowsFileSystemProvider.implDelete(WindowsFileSystemProvider.java:269)
[junit4] > at
sun.nio.fs.AbstractFileSystemProvider.delete(AbstractFileSystemProvider.java:103)
[junit4] > at java.nio.file.Files.delete(Files.java:1126)
[junit4] > at
org.apache.solr.handler.TestCSVLoader.tearDown(TestCSVLoader.java:64)
[junit4] > at java.lang.Thread.run(Thread.java:748)
I am not sure behind the reasoning of forcefully preventing closing of
ServletInputStreams - but IMHO, the better way to handle this is (we had this
discussion already a while ago):
Wrap the ServletInput/ServletOutput streams with a CloseShieldXxxxStream and
only pass this one around. This allows that code anywhere in solr is free to
close any streams correctly (which it should), but the ServletStreams are kept
open by the shield.
The issue we have no is that ContentStreams are not necessarily (like in
BlobUploadHandler, CSVHandler) coming from the servlet streams. If it is a file
or a uploaded file, then we HAVE to close the stream.
The reason behind everything here is a bug in Jetty - in contrast to Tomcat and
all other servlet containers, it closes the socket after you close the servlet
streams. This is a bug - sorry! Jetty should prevent closing the stream!
Please revert the current commit. I can help in solving this correctly -
unfortunately I am on travel next week.
> Do not close any servlet streams and improve our servlet stream closing
> prevention code for users and devs.
> -----------------------------------------------------------------------------------------------------------
>
> Key: SOLR-12290
> URL: https://issues.apache.org/jira/browse/SOLR-12290
> Project: Solr
> Issue Type: Bug
> Security Level: Public(Default Security Level. Issues are Public)
> Reporter: Mark Miller
> Assignee: Mark Miller
> Priority: Major
> Fix For: 7.4, master (8.0)
>
> Attachments: SOLR-12290.patch, SOLR-12290.patch, SOLR-12290.patch,
> SOLR-12290.patch
>
>
> Original Summary:
> When you fetch a file for replication we close the request output stream
> after writing the file which ruins the connection for reuse.
> We can't close response output streams, we need to reuse these connections.
> If we do close them, clients are hit with connection problems when they try
> and reuse the connection from their pool.
> New Summary:
> At some point the above was addressed during refactoring. We should remove
> these neutered closes and review our close shield code.
> If you are here to track down why this is done:
> Connection reuse requires that we read all streams and do not close them -
> instead the container itself must manage request and response streams. If we
> allow them to be closed, not only do we lose some connection reuse, but we
> can cause spurious client errors that can cause expensive recoveries for no
> reason. The spec allows us to count on the container to manage streams. It's
> our job simply to not close them and to always read them fully, from client
> and server.
> Java itself can help with always reading the streams fully up to some small
> default amount of unread stream slack, but that is very dangerous to count
> on, so we always manually eat up anything on the streams our normal logic
> ends up not reading for whatever reason.
> We also cannot call abort without ruining the connection or sendError. These
> should be options of very last resort (requiring a blood sacrifice) or when
> shutting down.
>
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscr...@lucene.apache.org
For additional commands, e-mail: dev-h...@lucene.apache.org
