[jira] [Issue Comment Deleted] (OMID-90) Reducing begin/commit latency by distributing the write to the commit table

["Lianghong Xu (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/OMID-90?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Lianghong Xu updated OMID-90:
-----------------------------
    Comment: was deleted

(was: Hi [~yonigo],

We were using ycsb 0.15.0-SNAPSHOT. The experiment setup is as follow:
Hardware:
12 Omid clients: i3.2xlarge instances. 
2 TSO servers: HA setup, c3.8xlarge instances
20 HBase regionservers: i3.2xlarge instances.

YCSB:
1M records, 80/20 read/write ratio. uniform dist. filedcount = 4. For original Omid, we used 

numConcurrentCTWriters: 8
batchSizePerCTWriter: 50

The max throughput we could obtain for both original and OmidLL is around 50-65K. We tried various tuning ourselves but couldn't seem to get it higher. 


Thanks,
Lianghong)

> Reducing begin/commit latency by distributing the write to the commit table
> ---------------------------------------------------------------------------
>
>                 Key: OMID-90
>                 URL: https://issues.apache.org/jira/browse/OMID-90
>             Project: Apache Omid
>          Issue Type: New Feature
>            Reporter: Ohad Shacham
>            Assignee: Yonatan Gottesman
>            Priority: Major
>         Attachments: OmidCloud-VLDB.pdf
>
>
> Today, Omid's commits are done by the transaction manager. In order to efficiently write to the commit table, the transaction manager batches these writes. This optimization, even thought reduces the write time to HBase, significantly increases the begin and commit latency. The commit latency increases since a commit operation returns only after its commit timestamp was persisted in the commit table. And the begin latency increases since begin returns a transaction id that is also used by the transaction to identify its snapshot and therefore, begin returns only after all commits with commit id smaller than the begin id was persisted in the commit table. This is crucial, since a snapshot change during a transaction run may violate snapshot isolation. 
>  
> The idea of this feature is to distribute the commit by moving the write to the commit table from the server to the client. The transaction manager does conflict analysis and returns a commit timestamp. While the client atomically persists this commit in the commit table.
> This significantly reduces the begin and commit latency, since batching is not required anymore. A begin operation can immediately returns and a commit operation returns after conflict detection. 
> This can introduce snapshot isolation violation since a slow client can commit and change other transaction's snapsho. Therefore, we use an invalidation technique which is similar to the one Omid uses today to maintain snapshot isolation in high availability mode.



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