[
https://issues.apache.org/jira/browse/ACCUMULO-132?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13147847#comment-13147847
]
Keith Turner commented on ACCUMULO-132:
---------------------------------------
If the problem is slow tablet servers, then another way to deal with this is to
detect and decommission slow tablet servers.
If the problem is a slow datanode, many of the proposed solutions may still
result in slow reads. In this case a solution may be to detect and
decommission slow data nodes.
When decommissioning datanodes and tablet servers, would need to ensure the
whole cluster is not decommissioned :)
Another source of slow reads I have seen is at the network level. Once I
observed an entire rack running slow. After manually removing all nodes in
this rack, everything sped up.
> read-only access to existing tables for speculative execution
> -------------------------------------------------------------
>
> Key: ACCUMULO-132
> URL: https://issues.apache.org/jira/browse/ACCUMULO-132
> Project: Accumulo
> Issue Type: New Feature
> Components: client
> Affects Versions: 1.5.0
> Reporter: Eric Newton
> Assignee: Eric Newton
>
> An accumulo user has a need to perform a map/reduce over an entire table.
> Right now, if a tablet server is running on a machine that has become slow
> due to failing hardware (it happens... a lot), the mapper using that server
> also runs slowly. If we allow speculative execution, the second map just
> goes back to the same tablet server, doubling the load on our already slow
> server. However, the data behind the tablet server is sitting there in
> multiple locations in HDFS. The user can live without pulling the most
> recent updates in the tablet server's memory.
> Possible solutions:
> # provide read-only hosting of a tablet from another node, possibly on demand
> # convenient way of re-creating the tablet server iterator stack and reading
> the files directly
> # clone the table/tablet (has the benefit that a reference is maintained in
> the metadata table)
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
