MaxGekk opened a new pull request #31819: URL: https://github.com/apache/spark/pull/31819
### What changes were proposed in this pull request? In non-ANSI mode, casting float to timestamp has different implementation for codegen on and off. Codegen on: 1. Multiply float input by MICROS_PER_SECOND 2. Cast resulting float value to long Codegen off: 1. CAST float input to double input 2. Multiply double input by MICROS_PER_SECOND 3. Cast resulting double value to long In the PR, I propose to align non-codegen code, and cast input float to double. ### Why are the changes needed? This fixes the issue which is demonstrated by the code: ```sql spark-sql> CREATE TEMP VIEW v1 AS SELECT 16777215.0f AS f; spark-sql> SELECT * FROM v1; 1.6777215E7 spark-sql> SELECT CAST(f AS TIMESTAMP) FROM v1; 1970-07-14 07:20:15 spark-sql> CACHE TABLE v1; spark-sql> SELECT * FROM v1; 1.6777215E7 spark-sql> SELECT CAST(f AS TIMESTAMP) FROM v1; 1970-07-14 07:20:14.951424 ``` The result from the cached view **1970-07-14 07:20:14.951424** is different from un-cached view **1970-07-14 07:20:15**. ### Does this PR introduce _any_ user-facing change? Yes. After the changes, the example above outputs the same timestamp for the cached view: ```sql spark-sql> CACHE TABLE v1; spark-sql> SELECT * FROM v1; 1.6777215E7 spark-sql> SELECT CAST(f AS TIMESTAMP) FROM v1; 1970-07-14 07:20:15 ``` ### How was this patch tested? By running new test: ``` $ build/sbt "test:testOnly *CastSuite" ``` ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
