[
https://issues.apache.org/jira/browse/HBASE-17998?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16001881#comment-16001881
]
Karan Mehta commented on HBASE-17998:
-------------------------------------
I did not find really feasible for this thing to happen because of following
reasons.
1. Calling {{getSerializedSize()}} on a Protobuf to estimate its size will
require to traverse over the complete protobuf data which we don't wanna do in
the first place.
2. Size of Puts can be estimated based on the number of bytes received for the
RPC Request. This information can be passed around with {{HBaseRpcController}}
class. However, it is difficult to estimate the size of only {{Puts}} in case
of a {{MultiRequest}} since we just have a total buffer size which may include
{{Gets}} and {{Scans}}. I am not sure if we should really be doing code changes
for just this support.
Rather than randomly estimating the size, a slightly better approach might be
to use a moving average based on the past requests that have been seen as
suggested by [~vincentpoon]. Lets discuss if there are any other ways to
accomplish this task.
> Improve HBase RPC write throttling size estimation
> --------------------------------------------------
>
> Key: HBASE-17998
> URL: https://issues.apache.org/jira/browse/HBASE-17998
> Project: HBase
> Issue Type: Improvement
> Reporter: Karan Mehta
> Assignee: Karan Mehta
>
> Currently when RPC throttling, the size of each put is estimated using a
> hardcoded value 100 bytes. This can be improved by using the protobuf size as
> an estimate, without having to deserialize or do a big refactoring.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
