[
https://issues.apache.org/jira/browse/HBASE-22269?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16834752#comment-16834752
]
Zheng Hu edited comment on HBASE-22269 at 5/7/19 1:26 PM:
----------------------------------------------------------
Consider the case: we have 3 blocks, the acess order is: block1, block2,
block3. so the accessCounter of them will be : (block1, 1), (block2, 2),
(block3, 3). an RPC path is still reference to block1, while block2 and
block3 has no RPC path now. Should we evict the block1 ?
We cannot evict block1, because block1 is still being referenced and future RPC
will have greater possibility refer to block1 instead of block2 or block3. so
when considering eviction, seems we cannot just release the block from oldest
to newst based on accessCounter regardless of the RPC path reference.
So I plan to close this issue as *Won't fix* , FYI [~Apache9], [~anoop.hbase],
[~ram_krish]. How do you guys think ?
was (Author: openinx):
Consider the case: we have 3 blocks, the acess order is: block1, block2,
block3. so the accessCounter of them will be : (block1, 1), (block2, 2),
(block3, 3). an RPC path is still reference to block1, while block2 and
block3 has no RPC path now. Should we evict the block1 ?
We cannot evict block1, because block1 is still being referenced and future RPC
will have greater possibility refer to block1 instead of block2 or block3. so
when considering eviction, seems we cannot just release the bucket from oldest
to newst based on accessCounter regardless of the RPC path reference.
So I plan to close this issue as *Won't fix* , FYI [~Apache9], [~anoop.hbase],
[~ram_krish]. How do you guys think ?
> Consider simplifying the logic of BucketCache eviction.
> -------------------------------------------------------
>
> Key: HBASE-22269
> URL: https://issues.apache.org/jira/browse/HBASE-22269
> Project: HBase
> Issue Type: Sub-task
> Reporter: Zheng Hu
> Priority: Major
>
> As discussed in review board: https://reviews.apache.org/r/70465 . [~Apache9]
> has an comment:
> bq. I think with the new reference counted framework, we do not need to treat
> rpc reference specially? Just release the bucket from oldest to newest, until
> we can find enough free space? We could know if the space has been freed from
> the return value of release ? Can be a follow on issue, maybe.
> Now, we'll choose those non-RPC refered block to mark as evicted, maybe can
> simplify the logic here , just as [~Apache9] said.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
