[
https://issues.apache.org/jira/browse/IMPALA-8685?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16869688#comment-16869688
]
Todd Lipcon commented on IMPALA-8685:
-------------------------------------
I think we need to be careful not to do anything that is surprising to end
user/admin expectations. In particular I think people might have the following
assumptions:
*If my working set is N GB of data, if I allocate N GB of cache across the
cluster I should be able to get 100% cache hit ratio. Some small fixed
percentage of overhead would be acceptable (eg if we say we recommend 120G of
cache for 100G fo data) but I think asking for 2x or 3x of working set is
unacceptable.*
This is my big fear about any default configuration other than 1 for
NUM_REMOTE_EXECUTOR_CANDIDATES.
*Given a cluster with N nodes (with some per-node cache configuration), a
cluster with N+1 nodes should have same or better cache rate*
This assumption would break if we go with something like {{min(3, half of
cluster size)}}, right? eg assuming we round down for "half of cluster size", a
3-node cluster would use 1 cache replica, but a 4-node cluster would use 2
cache replicas. Thus the cacheable working set for 3 nodes (100GB cache each)
would be 300GB whereas the cacheable working set for 4 nodes (100GB each) would
be 200GB, resulting in a reduced cache hit rate from an increased resource
allocation.
> Evaluate default configuration of NUM_REMOTE_EXECUTOR_CANDIDATES
> ----------------------------------------------------------------
>
> Key: IMPALA-8685
> URL: https://issues.apache.org/jira/browse/IMPALA-8685
> Project: IMPALA
> Issue Type: Improvement
> Components: Backend
> Reporter: Michael Ho
> Priority: Critical
>
> The query option {{NUM_REMOTE_EXECUTOR_CANDIDATES}} is set to 3 by default.
> This means that there are potentially 3 different executors which can process
> a remote scan range. Over time, the data of a given remote scan range will be
> spread across these 3 executors. My understanding of why this is not set to 1
> is to avoid hot spots in pathological cases. On the other hand, this may mean
> that we may not maximize the utilization of the file handle cache and data
> cache. Also, for small clusters (e.g. a 3 node cluster), the default value
> may render deterministic remote scan range scheduling ineffective. We may
> want to re-evaluate the default value of {{NUM_REMOTE_EXECUTOR_CANDIDATES}}.
> One idea is to set it to min(3, half of cluster size) so it works okay with
> small cluster, which may be rather common for demo purposes. There may also
> be other criteria for evaluating the default value.
> cc'ing [~joemcdonnell], [~tlipcon] and [~drorke]
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-all-unsubscr...@impala.apache.org
For additional commands, e-mail: issues-all-h...@impala.apache.org
