Snowflake 查询优化器是否尊重 CTE？

Question

If I write a SQL statement that includes CTEs, will the query optimiser always retain those CTEs as discrete statements to be optimised individually or can it merge those CTEs with other parts of the overall SQL if it calculates the resulting SQL will be quicker to execute?

This question was triggered by a question another user asked. They were using a sequence generator in a CTE; when the CTE SQL was run in isolation it always produced 12 consecutive numbers, as expected. However, when run in the CTE as part of a much larger SQL statement it was missing numbers ie it wasn't producing consecutive values.

This is a known issue/behaviour with large datasets but as there were only 12 values it shouldn't have been a problem - but the fact that it was suggests that the CTE was not being run as written, with the 12 record resultset then joined to the other tables, but instead the query optimiser had re-written the overall query and had merged the CTE logic with other parts of the SQL statement and so was prodcuing a much larger dataset.

Answer 1

Snowflake does not provide a lot of explanation of how it optimizes queries.

I can say that in general, there are two ways to handle CTEs:

• Materialize the CTE so it is run once and then the materialized version is read.
• Incorporate the CTE logic into the the rest of the query and optimize as part of the query.

I would actually expect Snowflake to do both , choosing the better execution plan, because it is a modern database that has learned from decades of optimization experience. Either method might be better under some circumstances.

All that said, if the data returned by the code changes as you describe it, then there is a bug. The point of optimization is not to change the semantics (meaning) of the code. The point is to return the same results but using different underlying algorithms.