[
https://issues.apache.org/jira/browse/OAK-9158?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Marcel Reutegger resolved OAK-9158.
-----------------------------------
Fix Version/s: 1.40.0
Resolution: Fixed
Committed changes to trunk: http://svn.apache.org/r1886954
Thank you for reporting this issue and providing a patch.
> Performance issue due to AbstractDocumentNodeState#equals
> ---------------------------------------------------------
>
> Key: OAK-9158
> URL: https://issues.apache.org/jira/browse/OAK-9158
> Project: Jackrabbit Oak
> Issue Type: Bug
> Components: documentmk
> Affects Versions: 1.10.8, 1.22.3
> Reporter: Alexander Lüders
> Assignee: Marcel Reutegger
> Priority: Major
> Fix For: 1.40.0
>
> Attachments: inheritance.png, patch.txt, stacktrace.txt
>
>
> TL;DR: It seems like OAK-7401 broke the symmetry contract of
> AbstractDocumentNodeState#equals method. This may lead to lots of expensive
> queries during commit.
> ---
> We have a MongoDB as a backend of a Oak repository where we see hundreds of
> expensive queries being executed during a rather simple operation.
> It happens when we are saving a session in which we added a child node to a
> node already containing a very large number of children.
> This has not been the case with Jackrabbit Oak 1.4.8 and 1.4.26 but it is an
> issue with 1.10.8 and 1.22.3.
> Newer versions have not been tested yet but looking at the source we believe
> that they will show that performance issue too.
> +What we know so far:+
>
> * The queries against the _MongoDocumentStore_ are triggered via
> _AbstractNodeState.equals_ (see attached stacktrace)
> * A TODO "inefficient unless there are very few child nodes" in that method
> indicates poor performance but this has been there since a long time and
> cannot explain the issue
> * OAK-7401 introduced the class _ModifiedDocumentNodeState_ class which is
> not inheriting _ModifiedNodeState_ or _AbstractDocumentNodeState_ (see
> attached class inheritance diagram)
> * _AbstractDocumentNodeState#equals_ is checking against instances of
> _ModifiedNodeState_ and _AbstractDocumentNodeState_ but not
> _ModifiedDocumentNodeState_.
> * _AbstractDocumentNodeState#equals_ triggers the "slow"
> _AbstractNodeState#equals_ as a last resort
>
> Guessing from a high-level perspective it looks like the introduction of
> _ModifiedDocumentNodeState_ broke the equals logic in class
> _AbstractDocumentNodeState_.
> We compiled a patch that solves the performance issue. All test and
> integration tests of oak-document-store also passed.
> See corresponding attachment.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
