[
https://issues.apache.org/jira/browse/IGNITE-12620?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17152566#comment-17152566
]
Alexey Goncharuk commented on IGNITE-12620:
-------------------------------------------
[~gvvinblade] Since trait propagation logic was reworked, can you provide a
high-level logic that is executed in each {{derive}} method (the high-level
invariants that are not dependent on the implementation details) in javadoc?
Currently it's a bit hard to follow the target behavior.
> Calcite integration. Index Nested Loop Join/Hash Join
> -----------------------------------------------------
>
> Key: IGNITE-12620
> URL: https://issues.apache.org/jira/browse/IGNITE-12620
> Project: Ignite
> Issue Type: New Feature
> Reporter: Igor Seliverstov
> Assignee: Igor Seliverstov
> Priority: Major
> Time Spent: 10m
> Remaining Estimate: 0h
>
> We may implement the feature the next way:
> # For each row from left node consume whole dataset from right node
> # Pass join condition as an argument of request() method of the right node
> # In case a data source at right is an index scan
> ## If there is no cursor opened - create a new cursor using bounds from
> request
> ## If there is an existing cursor - just check the cursor was opened using
> the same condition as passed one.
> ## After the right node signals EOD consume a next row from left and repeat
> from p 2.
> # In case a data source at right is a table scan with huge amount of rows
> ## If there is no cursor opened - create a new cursor ignoring passed
> condition
> ## If there is an existing cursor - just check the cursor was opened using
> the same condition as passed one.
> ## After the right node signals EOD consume a next row from left and repeat
> from p 2.
> # In case a data source at right is remote / is a table scan with small
> number of rows/ is a filter node with good enough selectivity
> ## Create a hash index for a data source
> ## If there is no cursor opened - create a new cursor using bounds from
> request
> ## If there is an existing cursor - just check the cursor was opened using
> the same condition as passed one.
> ## After the right node signals EOD consume a next row from left and repeat
> from p 2.
> Consider implementation specifics at optimization time, choose the cheapest
> variant
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
