[
https://issues.apache.org/jira/browse/OAK-4888?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15595262#comment-15595262
]
Thomas Mueller commented on OAK-4888:
-------------------------------------
> Will this catch queries with a suboptimal index that still needs a traversal
> of the index
No, not yet. Currently there cost of indexes is not taken into account. It
could be, that's true, but I "hope" that we only have indexes that are
selective (for example, don't have a nodetype index for very common nodetypes).
As for whether or not this is a problem, I think we need more data (examples).
> say ordering by an unindexed property
If the result is small, then ordering shouldn't be a problem. And I think the
vast majority of queries have a small result set, so that wouldn't be a
problem. So the question would be, do we expect _every_ query can have a small
or large result.
> I don't know why a missing index can ever be considered ok
The question is, do you consider a path restriction to be an index or not. But
in general, creating an index has an overhead which is larger than running a
query, so if you run a query only once, then it's better to not create an index.
> osgi configuration
Long term, I would prefer if there is no config option at all (as that's yet
another knob that increases the surface area, and can be misused...). I was
hoping that we can make "fail" the default at some point, and then we can even
remove the (undocumented) system property.
> we may need a way to allow cancel of a query
Yes, we should have a "management level" tool for that. Right now, it's
possible to cancel all long running queries by setting the "read limit" in the
query engine settings (JMX bean) by the way.
> Warn or fail queries above a configurable cost value
> ----------------------------------------------------
>
> Key: OAK-4888
> URL: https://issues.apache.org/jira/browse/OAK-4888
> Project: Jackrabbit Oak
> Issue Type: Improvement
> Components: query
> Reporter: Alexander Klimetschek
> Assignee: Thomas Mueller
> Labels: docs-impacting
> Fix For: 1.5.13
>
>
> *Problem:* It's easy to run into performance problems with queries that are
> not backed by an index or miss the right one. Developers writing these
> queries typically do not have the real large production data, and thus don't
> see that a query would scale badly, and would not see any traversal warnings,
> as these only happen with a large number of results.
> *Proposal:* Oak's query engine already calculates a cost estimate to make a
> decision which index to use, or even if there is any at all. This cost
> estimate could be used to find out if a query is not supported by an index or
> with one that is not suitable enough (e.g. ordering by property that is not
> indexed)
> If a query is above a certain cost value, a big warning could be put out or
> even the query could be made to fail (maybe a per query option, that you
> might want to have to "fail" by default to ensure people are not overlooking
> the problem). The cost limit should be configurable, as it might depend on
> the hardware power.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
