[
https://issues.apache.org/jira/browse/IGNITE-7015?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16294317#comment-16294317
]
Roman Kondakov commented on IGNITE-7015:
----------------------------------------
[~vozerov], I optimized my patch as you wrote in a comment above. A set of the
indexed columns was added to {{GridH2Table}}. This set is changed on each index
add/remove action. Temporary indexes are considered as normal indexes in this
case. All comparison logic moved to {{GridH2Table}}.
As for {{GridQueryProcessor.store}} invocation - I went through the code
execution in case of in-place update is possible and made sure that
{{GridQueryProcessor.store}} is not executed in this case because in the method
{{GridCacheMapEntry.AtomicCacheUpdateClosure#update}} we can see this snippet:
{code:java}
treeOp = oldRow != null && oldRow.link() == newRow.link() ?
IgniteTree.OperationType.NOOP :
IgniteTree.OperationType.PUT;
{code}
which is responsible for a choosing further tree operation after the closure
update. And if rows links are equal (which is true for in-place update), the
{{NOOP}} is chosen and therefore methods {{CacheDataStoreImpl#finishUpdate}}
and {{GridQueryProcessor#store}} are not called in
{{CacheDataStoreImpl#invoke}}.
TC tests are OK. Please review.
> SQL: index should be updated only when relevant values changed
> --------------------------------------------------------------
>
> Key: IGNITE-7015
> URL: https://issues.apache.org/jira/browse/IGNITE-7015
> Project: Ignite
> Issue Type: Task
> Components: sql
> Reporter: Vladimir Ozerov
> Assignee: Roman Kondakov
> Labels: iep-1, performance
> Fix For: 2.4
>
>
> See {{GridH2Table.update}} method. Whenever value is updated, we propagate it
> to all indexes. Consider the following case:
> 1) Old row is not null, so this is "update", not "create".
> 2) Link hasn't changed
> 3) Indexed fields haven't changed
> If all conditions are met, we can skip index update completely, as state
> before and after will be the same. This is especially important when
> persistence is enabled because currently we generate unnecessary dirty pages
> what increases IO pressure.
> Suggested fix:
> 1) Iterate over index columns, skipping key and affinity columns (as they are
> guaranteed to be the same);
> 2) Compare relevant index columns of both old and new rows
> 3) If all columns are equal, do nothing.
> Fields should be read through {{GridH2KeyValueRowOnheap#getValue}}, because
> in this case we will re-use value cache transparently.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
