Why redshift table structure getting changed while loading with glue job?

The reason for that is that for efficiency the table is dumped to the s3 temporary path specified as CSV, this means some specific types are lost and it's possible that some columns have characters that make it incorrectly parse as more columns that it really has.

Glue 4.0 has a new improved connector to Redshift, can you still reproduce it in that version? If that's the case, please open a support case or indicate here how to reproduce the issue with the multiple column (e.g. what character is confusing the CSV parsing).