[jira] [Commented] (HUDI-3834) Evaluate MT Column Stats Performance

["Alexey Kudinkin (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HUDI-3834?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17519769#comment-17519769 ] 

Alexey Kudinkin commented on HUDI-3834:
---------------------------------------

Original runtime of the query reading whole CSI from MT (takes about 60s):

!Screen Shot 2022-04-07 at 7.48.04 PM.png|width=610,height=763!

 

 

Taking up some basic CPU sampling profile, reveals that most of the time *~55%* is spent in `SpecificData.getForSchema`:

!Screen Shot 2022-04-07 at 7.46.01 PM.png|width=1321,height=771!

> Evaluate MT Column Stats Performance 
> -------------------------------------
>
>                 Key: HUDI-3834
>                 URL: https://issues.apache.org/jira/browse/HUDI-3834
>             Project: Apache Hudi
>          Issue Type: Bug
>            Reporter: Alexey Kudinkin
>            Assignee: Alexey Kudinkin
>            Priority: Blocker
>             Fix For: 0.11.0
>
>         Attachments: Screen Shot 2022-04-07 at 7.46.01 PM.png, Screen Shot 2022-04-07 at 7.48.04 PM.png
>
>
> h3. *UPDATE*
> --------
> *TL;DR* After identifying the bottlenecks as Avro 1.10 regression in generated Builder classes, relying on `SpecificData.getForSchema` call that loads corresponding model's class using reflection, we're able to bring down in a similar setting the runtime of reading for Column Stats Index from MT containing 800k records from *60s* to *3s* (20x)
>  
> Follow the comments for updates on that.
> --------
> Previously, while evaluating Data Skipping runtime in EMR setting, it was measured that reading of Column Stats Index of about ~800k records takes about *60s* (~10ms / record).
>  
> Given that the total size of the log-files involved is ~60Mb, seems like there are some performance bottlenecks that we should investigate before 0.11 release



--
This message was sent by Atlassian Jira
(v8.20.1#820001)