[
https://issues.apache.org/jira/browse/YARN-3553?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14523757#comment-14523757
]
Craig Welch commented on YARN-3553:
-----------------------------------
[~xinxianyin], this pattern is important for these implementations for
efficiency reasons - the frequency with which changes occur which effect
ordering is much less than the frequency with which the applications must be
available in the proper order for allocation (esp. when allocating on
heartbeat, which is typical..). By iteratively resorting individual elements
only when needed we avoid frequent resorting of all applications in the queue,
which can be quite expensive with the frequency it would occur. Values are
cached with a pretty simple update lifecycle to avoid the issues you are
concerned about. Finally, this is an implementation specific choice, other
implementations of ordering policies are free to use other data structures /
sorting frequency, although the concern wrt efficiency which this approach
avoids would apply to any non-iterative approaches.
> TreeSet is not a nice container for organizing schedulableEntities.
> -------------------------------------------------------------------
>
> Key: YARN-3553
> URL: https://issues.apache.org/jira/browse/YARN-3553
> Project: Hadoop YARN
> Issue Type: Wish
> Components: scheduler
> Reporter: Xianyin Xin
>
> For TreeSet, element is identified by comparator, not the object reference.
> If any *attributes that used for comparing two elements* of an specific
> element is modified by other methods, the TreeSet will be in an un-sorted
> state, and cannot become sorted forever except that we reconstruct another
> TreeSet with the elements. To avoid this, one must be *very careful* when
> they try to modify the attributes (such as increase or decrease the used
> capacity of a schedulabeEntity) of an object.
> An example in AbstractComparatorOrderingPolicy.java, Line63,
> {code}
> protected void reorderSchedulableEntity(S schedulableEntity) {
> //remove, update comparable data, and reinsert to update position in order
> schedulableEntities.remove(schedulableEntity);
> updateSchedulingResourceUsage(
> schedulableEntity.getSchedulingResourceUsage());
> schedulableEntities.add(schedulableEntity);
> }
> {code}
> This method tries to remove the schedulableEntity first and then reinsert it
> so as to reorder the set. However, the changes of the schedulableEntity
> should be done in the middle of the above two operations. But the comparator
> of the class is not clear, so we don't know which attributes of the
> schedulableEntity was changed. If we changed the schedulableEntity outside
> the method and then inform the orderingPolicy that we made such a change, the
> operation "schedulableEntities.remove(schedulableEntity)" would not work
> correctly since the element of a TreeSet is identified by comparator. Any
> implement class of this abstract class should overwrite this method, but few
> does. Another choice is that we make modification of a schedulableEntity
> manually, but we mustn't forget to reorder the set when we do so and must
> remember the order: remove, modify the attributes(used for comparing),
> insert, or use an iterator to mark the schedulableEntity so that we can
> remove and reinsert it correctly.
> YARN-897 is an example that we fell into the trap. If the comparator become
> complex in future, e.g., we consider other types of resources in comparator,
> such traps will be more and disperse anywhere, which makes it easy to let a
> TreeSet become a un-sorted state.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
