[
https://issues.apache.org/jira/browse/DRILL-4446?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15180050#comment-15180050
]
ASF GitHub Bot commented on DRILL-4446:
---------------------------------------
Github user vkorukanti commented on the pull request:
https://github.com/apache/drill/pull/403#issuecomment-192332765
Patch update.
+ Address review comments including unittest for
HardAffinityFragmentParallelizer
+ Fix Limit0 issue when one of the GroupScans in limit0 query has hard
assignment requirements (in which case we can't run the entire query in one
fragment). @sudheeshkatkam Can you review this change for any perf
implications? I tested limit0 queries on parquet table containing more than 1k
files. I didn't see any regressions.
> Improve current fragment parallelization module
> -----------------------------------------------
>
> Key: DRILL-4446
> URL: https://issues.apache.org/jira/browse/DRILL-4446
> Project: Apache Drill
> Issue Type: New Feature
> Affects Versions: 1.5.0
> Reporter: Venki Korukanti
> Assignee: Venki Korukanti
> Fix For: 1.6.0
>
>
> Current fragment parallelizer {{SimpleParallelizer.java}} can’t handle
> correctly the case where an operator has mandatory scheduling requirement for
> a set of DrillbitEndpoints and affinity for each DrillbitEndpoint (i.e how
> much portion of the total tasks to be scheduled on each DrillbitEndpoint). It
> assumes that scheduling requirements are soft (except one case where Mux and
> DeMux case where mandatory parallelization requirement of 1 unit).
> An example is:
> Cluster has 3 nodes running Drillbits and storage service on each. Data for a
> table is only present at storage services in two nodes. So a GroupScan needs
> to be scheduled on these two nodes in order to read the data. Storage service
> doesn't support (or costly) reading data from remote node.
> Inserting the mandatory scheduling requirements within existing
> SimpleParallelizer is not sufficient as you may end up with a plan that has a
> fragment with two GroupScans each having its own hard parallelization
> requirements.
> Proposal is:
> Add a property to each operator which tells what parallelization
> implementation to use. Most operators don't have any particular strategy
> (such as Project or Filter), they depend on incoming operator. Current
> existing operators which have requirements (all existing GroupScans) default
> to current parallelizer {{SimpleParallelizer}}. {{Screen}} defaults to new
> mandatory assignment parallelizer. It is possible that PhysicalPlan generated
> can have a fragment with operators having different parallelization
> strategies. In that case an exchange is inserted in between operators where a
> change in parallelization strategy is required.
> Will send a detailed design doc.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
