[ 
https://issues.apache.org/jira/browse/HBASE-24632?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17163800#comment-17163800
 ] 

Michael Stack commented on HBASE-24632:
---------------------------------------

{quote}So for every WAL file, it will add a sub procedure in ProcedureExecutor 
and that in turn add entries for RemoteProcExecutor correct?
{quote}
[~anoop.hbase] yes.

 
{quote}So at this stage also these Sub proc executing threads will yield once 
the split proc is submitted to RemoteProcExecutor ?
{quote}
yes.... previous SCP would block occupying the PE worker thread until WAL split 
competed.

 
{quote}And there in that executor we have 128 threads by default.
{quote}
 

{color:#000080}It keeps the below number of threads running:{color}

 

{color:#000080}final int {color}numThreads = 
{color:#660e7a}conf{color}.getInt(MasterProcedureConstants.{color:#660e7a}MASTER_PROCEDURE_THREADS{color},
 Math.max(
 (cpus > {color:#0000ff}0 {color}? cpus / {color:#0000ff}4 {color}: 
{color:#0000ff}0{color}), 
MasterProcedureConstants.{color:#660e7a}DEFAULT_MIN_MASTER_PROCEDURE_THREADS{color}));

 

... where {color:#660e7a}DEFAULT_MIN_MASTER_PROCEDURE_THREADS is 16.
{color}

{color:#660e7a}... and internally, it will spin up to{color}

{color:#0000ff}10 {color}* numThreads;

 

... if it gets backed up.

 
{quote}Or will that also yield?
{quote}
It will yield.

 

Ok to commit? Thanks [~anoop.hbase]

 

 

 

> Enable procedure-based log splitting as default in hbase3
> ---------------------------------------------------------
>
>                 Key: HBASE-24632
>                 URL: https://issues.apache.org/jira/browse/HBASE-24632
>             Project: HBase
>          Issue Type: Sub-task
>          Components: wal
>            Reporter: Michael Stack
>            Priority: Major
>             Fix For: 3.0.0-alpha-1, 2.4.0
>
>
> Means changing this value in HConstants to false:
>    public static final boolean DEFAULT_HBASE_SPLIT_COORDINATED_BY_ZK = true;
> Should probably also deprecate the current zk distributed split too so we can 
> clear out those classes to.



--
This message was sent by Atlassian Jira
(v8.3.4#803005)

Reply via email to