[
https://issues.apache.org/jira/browse/CALCITE-4994?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17483473#comment-17483473
]
Julian Hyde commented on CALCITE-4994:
--------------------------------------
Yes, it would be a problem for the map in {{RelDataType}}. There is non-zero
chance that one thread will see the map is null, starts populating it, and
meanwhile another thread does the same thing. To do it properly you need to use
a {{volatile}} keyword, or use double-checked locking, or protect it with
{{Suppliers.memoize}}, and thoroughly test it, but we'd never know for sure.
Not worth the hassle.
I haven't benchmarked it, but I think that the last solution - that uses
{{MapBackedList}} - is better than all others. Memory usage is barely higher
than today, the map is used throughout the lifecycle, there's no mutable state,
and lookup performance is the best possible.
> SqlToRelConverter creates FieldMap for every Identifier Instead of Memoizing
> it
> -------------------------------------------------------------------------------
>
> Key: CALCITE-4994
> URL: https://issues.apache.org/jira/browse/CALCITE-4994
> Project: Calcite
> Issue Type: Improvement
> Components: core
> Reporter: Jay Narale
> Priority: Major
> Labels: pull-request-available
> Time Spent: 0.5h
> Remaining Estimate: 0h
>
> When converting from Sql To Rel, In SqlToRelConverter for every single
> instance of an identifier we create a new map in
> *_org.apache.calcite.sql2rel.SqlToRelConverter.Blackboard#lookupExp_*
>
> {code:java}
> final Map<String, Integer> fieldOffsets = new HashMap<>();
> for (RelDataTypeField f : resolve.rowType().getFieldList()) {
> if (!fieldOffsets.containsKey(f.getName())) {
> fieldOffsets.put(f.getName(), f.getIndex());
> }
> }
> final Map<String, Integer> map = ImmutableMap.copyOf(fieldOffsets);{code}
>
> So for a Sql Query
> {code:java}
> SELECT name, nation FROM customer{code}
> We would do the above operation twice.
> Memoization of this information will improve performance.
> In my database, I had observed that for a large table involving 1200 columns
> and a huge select having multiple expressions and operators, this part was a
> bottleneck.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
