[
https://issues.apache.org/jira/browse/HBASE-22417?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16840430#comment-16840430
]
Wellington Chevreuil commented on HBASE-22417:
----------------------------------------------
Thanks for your comments, [~xucang]!
{quote}In DeleteTableProcedure state machine, there is one state called
"DELETE_TABLE_UPDATE_DESC_CACHE".
After this patch, it becomes redundant. Do you think it makes sense to clean it
up there?
{quote}
Yeah, this whole stage would be redundant indeed. I think having the cache
cleaned on *deleteFromMeta* is safer and would make it more cohesive. There
would be less chance of inconsistent information between the cache and current
meta state in an event of DeleteTableProcedure stuck or crashed between these
two stages. Also, since *deleteFromMeta* is not private, callers can be unaware
of additional need to clean this cache. That's the case for
TruncateTableProcedure for example. In cases where it calls
_DeleteTableProcedure.deleteFromMeta_ on its *TRUNCATE_TABLE_REMOVE_FROM_META*
stage, then takes long time on *TRUNCATE_TABLE_CLEAR_FS_LAYOUT* due to a slow
FS, for example, this inconsistency between cache and meta can cause an 500
error on the Web UI.
{quote}And there is a method called "deleteTableDescriptorCache", should we
call this one instead?
{quote}
Yeah, haven't noticed that one before, definitely makes sense to reuse it.
Am checking on the test failures. Let me know if you agree with the above
comment, if so I can also apply it on a following patch version.
> DeleteTableProcedure.deleteFromMeta method should remove table from Master's
> table descriptors cache
> ----------------------------------------------------------------------------------------------------
>
> Key: HBASE-22417
> URL: https://issues.apache.org/jira/browse/HBASE-22417
> Project: HBase
> Issue Type: Bug
> Reporter: Wellington Chevreuil
> Assignee: Wellington Chevreuil
> Priority: Major
> Attachments: HBASE-22417.master.001.patch
>
>
> DeleteTableProcedure defines a static deleteFromMeta method that's currently
> used both by DeleteTableProcedure itself and TruncateTableProcedure.
> Sometimes, depending on the table size (and under slower, under performing
> FileSystems), truncation can take longer to complete
> *TRUNCATE_TABLE_CLEAR_FS_LAYOUT* stage, but the given table has already been
> deleted from meta on previous *TRUNCATE_TABLE_REMOVE_FROM_META* stage. In
> this case, features relying on Master's table descriptor's cache might
> wrongly try to reference this truncating table. Master Web UI, for example,
> would try to check this table state and end up showing a 500 error.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
