[
https://issues.apache.org/jira/browse/HDFS-14210?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16762020#comment-16762020
]
Íñigo Goiri commented on HDFS-14210:
------------------------------------
Let's unify the discussion with [~ayushtkn] in HDFS-14226 here.
So the main question is what should happen with a mount point with multiple
subcluster destinations.
There are 4 types at this point:
* HASH: it will distribute the files across the subclusters and this should be
a mapping of one file (and folder) to one subcluster. If a file is under a
folder (not the root of the mount point) it will go to the subcluster of the
folder.
* HASH_ALL: It will distribute the files across the subclusters and this should
a mapping of one file to one subcluster and folders should be in all
subclusters.
* RANDOM: This is usually for load balancing and all files and folders should
be the same in all subclusters.
* LOCAL: It tries to write to the local subcluster of the writer. It should
create folders everywhere and the files in just one subcluster.
* SPACE: It tries to write to the subcluster based on available space. It
creates the folders everywhere but the files in just one subcluster
Currently, this follows {{RouterClientProtocol#isPathAll()}} which is
ultimately based on {{MountTable#isAll()}} which specifies that the folders
should be in all location (i.e., HASH_ALL, RANDOM, and SPACE).
>From this, it looks like LOCAL is wrong as it should create the folders
>everywhere.
There are some inconsistencies here:
* We should have a distinction between folders in all subclusters (HASH_ALL,
RANDOM, LOCAL, and SPACE) and file in all subclusters (only RANDOM).
* SPACE and LOCAL don't have an order that can be precomputed so we need to
check in all of them (HASH_ALL and HASH we know where the file should be in the
common case).
* For files in one subcluster (all except RANDOM), we should just execute
commands where the file actually is.
> RBF: ModifyACL should work over all the destinations
> ----------------------------------------------------
>
> Key: HDFS-14210
> URL: https://issues.apache.org/jira/browse/HDFS-14210
> Project: Hadoop HDFS
> Issue Type: Sub-task
> Reporter: Shubham Dewan
> Assignee: Shubham Dewan
> Priority: Major
> Attachments: HDFS-14210-HDFS-13891.002.patch,
> HDFS-14210-HDFS-13891.003.patch, HDFS-14210.001.patch
>
>
> 1) A mount point with multiple destinations.
> 2) ./bin/hdfs dfs -setfacl -m user:abc:rwx /testacl
> 3) where /testacl => /test1, /test2
> 4) command works for only one destination.
> ACL should be set on both of the destinations.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
