Peter Somogyi commented on HBASE-17449:

I had a discussion with [~stack] and he mentioned that 
{{hbase.client.operation.timeout}} should limit the total time for the 
operation (e.g. get) but as I noticed it is not exactly how it works currently. 
What I've tested is the following:

hbase.client.operation.timeout=3000ms, hbase.rpc.timeout=1000ms, 
hbase.client.retries.number=10, injected delay for get calls=1500ms. Based on 
this the call should finish after 3 seconds due to operation timeout but in the 
test the call will be executed 11 times (original + 10 retries).

Do we have a bug here?

> Add explicit document on different timeout settings
> ---------------------------------------------------
>                 Key: HBASE-17449
>                 URL: https://issues.apache.org/jira/browse/HBASE-17449
>             Project: HBase
>          Issue Type: Improvement
>          Components: documentation
>            Reporter: Yu Li
>            Assignee: Yu Li
>            Priority: Critical
>             Fix For: 2.0.0
>         Attachments: 0001-Standalone.patch, HBASE-17449.master.002.patch, 
> WIP-HBASE-17499.master.001.patch
> Currently we have more than one timeout settings, mainly includes:
> * hbase.rpc.timeout
> * hbase.client.operation.timeout
> * hbase.client.scanner.timeout.period
> And in latest branch-1 or master branch code, we will have two other 
> properties:
> * hbase.rpc.read.timeout
> * hbase.rpc.write.timeout
> However, in current refguid we don't have explicit instruction on the 
> difference of these timeout settings (there're explanations for each 
> property, but no instruction on when to use which)
> In my understanding, for RPC layer timeout, or say each rpc call:
> * Scan (openScanner/next): controlled by hbase.client.scanner.timeout.period
> * Other operations:
>    1. For released versions: controlled by hbase.rpc.timeout
>    2. For 1.4+ versions: read operation controlled by hbase.rpc.read.timeout, 
> write operation controlled by hbase.rpc.write.timeout, or hbase.rpc.timeout 
> if the previous two are not set.
> And hbase.client.operation.timeout is a higher-level control counting retry 
> in, or say the overall control for one user call.
> After this JIRA, I hope when users ask questions like "What settings I should 
> use if I don't want to wait for more than 1 second for a single 
> put/get/scan.next call", we could give a neat answer.

This message was sent by Atlassian JIRA

Reply via email to