[jira] [Updated] (HBASE-16973) Revisiting default value for hbase.client.scanner.caching

["Yu Li (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/HBASE-16973?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Yu Li updated HBASE-16973:
--------------------------
    Attachment: Scan.next_p999.png

Let me reword my question: Currently we set default of scanner caching to Integer.MAX_VALUE to make *_the network fill the chunk size defined by hbase.client.scanner.max.result.size rather than be limited by a particular number of rows_*, but what if the result never fills the chunk size due to some customized filter? Are we requesting too much for customers to know what size the result will be when using filters and explicitly set the caching size instead of simply using the default value? Should we use a smaller value or even no caching as the default and let advanced users to set it for performance optimization?

Attached is a screenshot of the server side monitoring metrics of p999 of scan.next, which reduces from 1min+ to few seconds after customer setting the caching size from default to 10, FWIW.

> Revisiting default value for hbase.client.scanner.caching
> ---------------------------------------------------------
>
>                 Key: HBASE-16973
>                 URL: https://issues.apache.org/jira/browse/HBASE-16973
>             Project: HBase
>          Issue Type: Bug
>            Reporter: Yu Li
>            Assignee: Yu Li
>         Attachments: Scan.next_p999.png
>
>
> We are observing below logs for a long-running scan:
> {noformat}
> 2016-10-30 08:51:41,692 WARN  [B.defaultRpcServer.handler=50,queue=12,port=16020] ipc.RpcServer:
> (responseTooSlow-LongProcessTime): {"processingtimems":24329,
> "call":"Scan(org.apache.hadoop.hbase.protobuf.generated.ClientProtos$ScanRequest)",
> "client":"11.251.157.108:50415","scandetails":"table: ae_product_image region: ae_product_image,494:
> ,1476872321454.33171a04a683c4404717c43ea4eb8978.","param":"scanner_id: 5333521 number_of_rows: 2147483647
> close_scanner: false next_call_seq: 8 client_handles_partials: true client_handles_heartbeats: true",
> "starttimems":1477788677363,"queuetimems":0,"class":"HRegionServer","responsesize":818,"method":"Scan"}
> {noformat}
> From which we found the "number_of_rows" is as big as {{Integer.MAX_VALUE}}
> And we also observed a long filter list on the customized scan. After checking application code we confirmed that there's no {{Scan.setCaching}} or {{hbase.client.scanner.caching}} setting on client side, so it turns out using the default value the caching for Scan will be Integer.MAX_VALUE, which is really a big surprise.
> After checking code and commit history, I found it's HBASE-11544 which changes {{HConstants.DEFAULT_HBASE_CLIENT_SCANNER_CACHING}} from 100 to Integer.MAX_VALUE, and from the release note there I could see below notation:
> {noformat}
> Scan caching default has been changed to Integer.Max_Value 
> This value works together with the new maxResultSize value from HBASE-12976 (defaults to 2MB) 
> Results returned from server on basis of size rather than number of rows 
> Provides better use of network since row size varies amongst tables
> {noformat}
> And I'm afraid this lacks of consideration of the case of scan with filters, which may involve many rows but only return with a small result.
> What's more, we still have below comment/code in {{Scan.java}}
> {code}
>   /*
>    * -1 means no caching
>    */
>   private int caching = -1;
> {code}
> But actually the implementation does not follow (instead of no caching, we are caching {{Integer.MAX_VALUE}}...).
> So here I'd like to bring up two points:
> 1. Change back the default value of HConstants.DEFAULT_HBASE_CLIENT_SCANNER_CACHING to some small value like 128
> 2. Reenforce the semantic of "no caching"



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)