lucene-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Uwe Schindler (JIRA)" <>
Subject [jira] [Commented] (SOLR-6519) In trunk change Solr's DirectoryFactory.create method to take LockFactory
Date Wed, 17 Sep 2014 18:23:35 GMT


Uwe Schindler commented on SOLR-6519:

I know why this happens:
In the new impl (after removing the hacky) injectLockFactory, I only added the "useful" ones
to each DirectoryFactory:
- HDFS only supports "none" or "hdfs"
- StandardDirectoryFactory only supports "simple", "native", "none"
- RAMDirectoryFactory only supports "single"
- MockDirectoryFactory ignores the factory completely, it uses what test framework chooses

I did not know that there was some additional randomization! For me the tests passed - strange.
We have 2 opinions:
- Add back "single" to StandardDirectoryFactory (which is a bit strange), but Hoss mentioned
on IRC, if somebody has a separate directory for every solr instance. It is still risky!
- Change the randomization

> In trunk change Solr's DirectoryFactory.create method to take LockFactory
> -------------------------------------------------------------------------
>                 Key: SOLR-6519
>                 URL:
>             Project: Solr
>          Issue Type: Bug
>            Reporter: Uwe Schindler
>            Assignee: Uwe Schindler
>             Fix For: 5.0
>         Attachments: SOLR-6519.patch, SOLR-6519.patch
> Because of NIO2 changes and the corresponding workaround, it is impossible now to create
a Directory and "hope" that the lock factory directory is not created. Especially if you want
some non-standard lock factory, this blows up.
> The problem is: The lock dir is now created in ctor. As workaround I made all factories
set NoLockFactory initially through ctor (see SOLR-6518), but this is just a workaround for
incorrect API design.
> In fact the main problem is just stupid: Why does protected CachingDirectoryFactory.create()
not take the lock factory? I think its because of backwards compatibility, but with Solr 5.0
we can change this.
> In future we want to make the lock factory non-mutable in Directory, so this is an important
change. In addition, injectLockFactory looks horrible, this code is a h├Ąckidy-hick-hack!

This message was sent by Atlassian JIRA

To unsubscribe, e-mail:
For additional commands, e-mail:

View raw message