You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@ignite.apache.org by "Vladimir Ozerov (JIRA)" <ji...@apache.org> on 2016/03/11 19:16:38 UTC

[jira] [Created] (IGNITE-2810) IGFS: Stripe TRASH directory.

Vladimir Ozerov created IGNITE-2810:
---------------------------------------

             Summary: IGFS: Stripe TRASH directory.
                 Key: IGNITE-2810
                 URL: https://issues.apache.org/jira/browse/IGNITE-2810
             Project: Ignite
          Issue Type: Task
          Components: IGFS
    Affects Versions: 1.5.0.final
            Reporter: Vladimir Ozerov
            Priority: Critical
             Fix For: 1.7


*Problem*
Currently remove operation on IGFS performs soft-delete which essentially moves a file/dir to a special hidden "TRASH" directory. Later on, special worker inspects this directory and finally removes actual meta and data entities.

As *move* operation requires lock on parent directory, massive removes will lead to high contention on TRASH key, yielding poor performance due to no parallelization.

*Solution*
Let's stripe trash directory. That is, we will create several (e.g. 16) trash directories. When entity is to be removed, we will pick random trash and move entitty there.

Delete worker will have to inspect all these directories.

Note that each node will have to know maximum numbers of trash directories, so it knows how to list it with little or no coordination with other nodes. To achieve this we can expose "trashCount" property to {{FileSystemConfiguration}}. Normally user will not change it. Each node will set up a discovery listener which will track new nodes and will know maximum amount of trashes.



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