[jira] [Comment Edited] (HBASE-20003) WALLess HBase on Persistent Memory

["Andrew Purtell (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HBASE-20003?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16368381#comment-16368381 ] 

Andrew Purtell edited comment on HBASE-20003 at 2/17/18 10:09 PM:
------------------------------------------------------------------

{quote}The existing region replica feature could be used here and ensure the data written to memstores are synchronously replicated to the replicas and ensure strong consistency of the data. (pipeline model)
{quote}
Let me be more precise about what I meant by "pmem doesn't obviate the need for a WAL unless it is replicated itself among multiple servers.".I mean the availability of the data in pmem needs to match today's data availability with the WAL or there is an overall availability loss.

Synchronous replication of edits from one region replica to another is a WAL by another name, but instead of the edit stream being available to the entire cluster in a replayable form it is limited to the three servers participating in the region replication. When all replicas go down at once, we have lost the ability to resume service for the affected region(s) on the other available hosts, because nobody beyond those replicas has any of the data. On a 1000+ node cluster, if you happen to lose 3 of the servers at once (which is more likely than you'd like, but the reality of scale ops) there is a good chance some regions have become completely unavailable until one of those servers can be brought back online. That is different from today, where every single server in the cluster has access to region data and WAL data in HDFS and can host the affected region(s).

Perhaps the PoC doc quantifies the availability loss? I'd be interested in taking a look. I suppose a case could be made that in some ways this matches the availability model of HDFS's default block placement policy, although HDFS does active mitigation of replica loss via re-replication and blocks are more dispersed than region replicas so an analysis is nontrivial.


was (Author: apurtell):
{quote}The existing region replica feature could be used here and ensure the data written to memstores are synchronously replicated to the replicas and ensure strong consistency of the data. (pipeline model)
{quote}
Let me be more precise about what I meant by "pmem doesn't obviate the need for a WAL unless it is replicated itself among multiple servers.".I mean the availability of the data in pmem needs to match today's data availability with the WAL or there is an overall availability loss.

Synchronous replication of edits from one region replica to another is a WAL by another name, but instead of the edit stream being available to the entire cluster in a replayable form it is limited to the three servers participating in the region replication. When all replicas go down at once, we lose the ability to resume service for the affected region(s) on the other available hosts.

Perhaps the PoC doc quantifies the availability loss? I'd be interested in taking a look.

 

> WALLess HBase on Persistent Memory
> ----------------------------------
>
>                 Key: HBASE-20003
>                 URL: https://issues.apache.org/jira/browse/HBASE-20003
>             Project: HBase
>          Issue Type: New Feature
>            Reporter: Anoop Sam John
>            Assignee: Anoop Sam John
>            Priority: Major
>
> This JIRA aims to make use of persistent memory (pmem) technologies in HBase. 
> One such usage is to make the Memstore to reside on pmem. Making a persistent memstore would remove the need for WAL and paves way for a WALLess HBase. 
> The existing region replica feature could be used here and ensure the data written to memstores are synchronously replicated to the replicas and ensure strong consistency of the data. (pipeline model)
> Advantages :
> -Data Availability : Since the data across replicas are consistent (synchronously written) our data is always 100 % available.
> -Lower MTTR : It becomes easier/faster to switch over to the replicas on a primary region failure as there is no WAL replay involved. Building the memstore map data also is much faster than reading the WAL and replaying the WAL.
> -Possibility of bigger memstores : These pmems are designed to have more memory than DRAMs so it would also enable us to have bigger sized memstores which leads to lesser flushes/compaction IO. 
> -Removes the dependency of HDFS on the write path
> Initial PoC has been designed and developed. Testing is underway and we would publish the PoC results along with the design doc sooner. The PoC doc will talk about the design decisions, the libraries considered to work with these pmem devices, pros and cons of those libraries and the performance results.
> Note : Next gen memory technologies using 3DXPoint gives persistent memory feature. Such memory DIMMs are soon to appear in the market. The PoC is done around Intel's ApachePass (AEP)



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)