[ 
https://issues.apache.org/jira/browse/HBASE-1185?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12678731#action_12678731
 ] 

Jean-Daniel Cryans commented on HBASE-1185:
-------------------------------------------

Billy, I can confirm that it's scanner pre-fetching. In HRS :

{code}
public RowResult[] next(final long scannerId, int nbRows) throws IOException {
    checkOpen();
    requestCount.incrementAndGet();
    ArrayList<RowResult> resultSets = new ArrayList<RowResult>();
    try {
      String scannerName = String.valueOf(scannerId);
      InternalScanner s = scanners.get(scannerName);
      if (s == null) {
        throw new UnknownScannerException("Name: " + scannerName);
      }
      this.leases.renewLease(scannerName);
      for(int i = 0; i < nbRows; i++) {
...
{code}

We increment for the whole batch of rows. I take the blame since it's my 
modification. The fix will be easy.

> wrong request/sec in the gui reporting wrong
> --------------------------------------------
>
>                 Key: HBASE-1185
>                 URL: https://issues.apache.org/jira/browse/HBASE-1185
>             Project: Hadoop HBase
>          Issue Type: Bug
>          Components: client, regionserver
>    Affects Versions: 0.19.0
>         Environment: 0.19.0
>            Reporter: Billy Pearson
>            Priority: Minor
>             Fix For: 0.20.0
>
>
> I am seeing lower number of request in the masters gui then I have seen in 
> 0.18.0 while scanning.
> I thank part of it is we moved to report per sec request not per 3 secs so 
> the request should be 1/3 of the old numbers I was getting.
> hbase.client.scanner.caching is not the reason the request are under reported.
> I set hbase.client.scanner.caching = 1 and still get about 2K request a sec 
> in the gui
> but when the job is done I take records / job time and get 36,324/ records 
> /sec. So
> there must be some caching out side of the hbase.client.scanner.caching 
> making the
> request per sec lower then it should be. I know it running faster then 
> reported just thought
> it might give some new users the wrong impression that request/sec = 
> read/write /sec.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.

Reply via email to