Caizhi Weng created FLINK-18687:
-----------------------------------
Summary: ProjectionCodeGenerator#generateProjectionExpression should remove for loop optimization
Key: FLINK-18687
URL:
https://issues.apache.org/jira/browse/FLINK-18687 Project: Flink
Issue Type: Bug
Components: Table SQL / Runtime
Affects Versions: 1.11.0
Reporter: Caizhi Weng
Fix For: 1.12.0
If too many fields of the same type are projected, {{ProjectionCodeGenerator#generateProjectionExpression}} currently performs a "for loop optimization" which, instead of generating code separately for each field, they'll be squashed into a for loop.
However, if the indices of the fields with the same type are not continuous, this optimization will not write fields in index ascending order. This is not acceptable because {{BinaryWriter}}s expect the users to write fields in index ascending order (that is to say, we *have to* first write field 0, then field 1, then...), otherwise the variable length area of the two binary rows with same data might be different. Although we can use {{getXX}} methods of {{BinaryRow}} to get the fields correctly, states for streaming jobs compare state keys with binary bits, not with the contents of the keys. So we need to make sure the binary bits of the binary rows be the same if two rows contain the same data.
What's worse, as the current implementation of {{ProjectionCodeGenerator#generateProjectionExpression}} uses a scala {{HashMap}}, the key order of the map might be different on different workers; Even if the projection does not meet the condition to be optimized, it will still be affected by this bug.
What I suggest is to simply remove this optimization. Because if we still want this optimization, we have to make sure that the fields of the same type have continuous order, which is a very strict and rare condition.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
Locked
