That behavior is similar to a number of analytic databases. It's expensive to maintain constraints in large distributed datasets. Referential integrity checks are also not meaningful in denormalized fact tables. Redshift [1] and ClickHouse [2] work this way as well. If things like duplicates are an issue, you can remove them by choosing query sort orders carefully, for example.
[1] https://community.snowflake.com/s/question/0D50Z00007Ft37mSA...
[2] https://stackoverflow.com/questions/35372889/a-simpler-more-...