[
https://issues.apache.org/jira/browse/HBASE-15212?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15132787#comment-15132787
]
Enis Soztutar commented on HBASE-15212:
---------------------------------------
bq. You mean TODO from HBASE-15177?
Yes, let me fix the typo.
bq. So when we move to reading reqs into off heap BB, the size calc based on
Xmx become not relevant any more. Any thinking on this?
It still might be relevant. Even with DBB IPC buffers, in case of Put, the
edits will go to memstore until that is also DBB. This is just a heuristic to
protect the server based on its heap. Lets see whether the heuristic makes
sense or not (suggestions welcome):
|| heap || numHandlers || max req size ||
| 1GB | 30 | 3.4MB |
| 1GB | 60 | 1.7MB |
| 4GB | 30 | 13.5MB |
| 4GB | 60 | 6.7MB |
| 8GB | 60 | 13.5MB |
| 8GB | 100 | 8MB |
| 16GB | 100 | 16MB |
| 32GB | 100 | 32MB |
>From the above, the heuristic maybe a bit too aggressive. It assumes that
>every handler gets a max-sized request at the same time.
> RRCServer should enforce max request size
> ------------------------------------------
>
> Key: HBASE-15212
> URL: https://issues.apache.org/jira/browse/HBASE-15212
> Project: HBase
> Issue Type: Improvement
> Reporter: Enis Soztutar
> Assignee: Enis Soztutar
> Fix For: 2.0.0, 1.3.0
>
> Attachments: hbase-15212_v1.patch
>
>
> A TODO from HBASE-15198 was that we are not protecting the RPCServer in case
> an RPC request with a very large size is received. This might cause the
> server to go OOM because we are allocating the RPC serialization into a BB.
> Instead we should reject the RPC and close the connection.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
