[
https://issues.apache.org/jira/browse/HBASE-15866?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Andrew Purtell updated HBASE-15866:
-----------------------------------
Fix Version/s: 1.4.0
I've done some poking around with 0.98. The code is significantly different.
RegionServerCallable and RpcRetryingCaller implementations differ, especially
RegionServerCallable#prepare (in 0.98 we use separate callers with rpc settings
derived from connection not table to do region relocation, so we don't
correctly honor the table level settings) and RpcRetryingCaller#callWithRetry
(uses a thread local for setting rpc timeouts in 0.98, later versions don't).
Setting a RPC timeout in configuration works. Both connection and table pick it
up. Using only the setter on HTable doesn't. Only the callables created by the
table use the setting; others created by the HConnection do not.
I'm not saying it can't be done, but it's much easier to achieve this with
master and branch-1, and why not have another reason to move off of 0.98.
Updated fix versions accordingly, only targeting master and branch-1.
If someone feels differently about porting this back to 0.98, by all means open
another issue for backport and tackle it there.
> Split hbase.rpc.timeout into *.read.timeout and *.write.timeout
> ---------------------------------------------------------------
>
> Key: HBASE-15866
> URL: https://issues.apache.org/jira/browse/HBASE-15866
> Project: HBase
> Issue Type: Bug
> Affects Versions: 2.0.0
> Reporter: Andrew Purtell
> Assignee: Vivek Koppuru
> Labels: newbie, patch
> Fix For: 2.0.0, 1.4.0
>
> Attachments: HBASE-15866.patch, HBASE-15866.patch, HBASE-15866.patch,
> read-write-rpc-timeouts.patch, read-write-rpc.v1.patch
>
>
> We have a single tunable for the RPC timeout interval - hbase.rpc.timeout.
> This is fine for the general case but there are use cases where it would be
> advantageous to set two separate timeouts for reads (gets, scans, perhaps
> with significant server side filtering - although the new scanner heartbeat
> feature mitigates where available) and mutations (fail fast under tight SLA,
> resubmit or take mitigating action).
> I propose we refer to a configuration setting "hbase.rpc.read.timeout" when
> handling read operations and "hbase.rpc.write.timeout" when handling write
> operations. If those values are not set in the configuration, fall back to
> the value of "hbase.rpc.timeout" or its default.
> So for example in HTable instead of one global timeout for each RPC
> (rpcTimeout), there would be a readRpcTimeout and writeRpcTimeout also set up
> in HTable#finishSetup. Then wherever we set up RPC with
> RpcRetryingCallerFactory#newCaller(int rpcTimeout) we pass in the read or
> write timeout depending on what the op is.
> In general I don't like the idea of adding configuration parameters to our
> already heavyweight set, but I think the inability to control timeouts
> separately for reads and writes is an operational deficit.
> See also PHOENIX-2916.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
