erratic "delayed" CTE evaluation? - snowflake-cloud-data-platform

Question

I observe a behaviour with CTEs which I did not expect (and seems inconsistent).
Not quite sure that it is correct...
Basically, through a CTE, I filter rows to avoid a particular problem, then use the result of that CTE to perform calculations that would break on the problematic rows which I thought I eliminated in my CTE...
Take a simple table with a varchar column that often has a number in it, but not always
CREATE TABLE MY_TABLE(ROW_ID INTEGER NOT NULL
, GOOD_ROW BOOLEAN NOT NULL
, SOME_VALUE VARCHAR NOT NULL);
INSERT INTO MY_TABLE(ROW_ID, GOOD_ROW, SOME_VALUE)
VALUES(1, TRUE, '1'), (2, TRUE, '2'), (3, FALSE, 'ABC');
I also create a small table with just numbers to join on
CREATE TABLE NUMBERS(NUMBER_ID INTEGER NOT NULL);
INSERT INTO NUMBERS(NUMBER_ID) VALUES(1), (2), (3);
Joining these two tables on SOME_VALUE results in an error because 'ABC' is not numeric and it appears that the JOIN is evaluated BEFORE the WHERE clause (BAD implications on performance here...)
SELECT *
FROM MY_TABLE
INNER JOIN NUMBERS ON NUMBERS.NUMBER_ID = TO_NUMBER(SOME_VALUE)
WHERE ROW_ID < 3; --> ERROR
So, I try to filter my first table through a CTE which only return rows for which SOME_VALUE is numeric
WITH ONLY_GOOD_ONES
AS (
SELECT SOME_VALUE
FROM MY_TABLE
WHERE GOOD_ROW = TRUE
)
SELECT *
FROM ONLY_GOOD_ONES;
Now, I would expect to be able to use the result of this CTE with SOME_VALUE being numeric.
WITH ONLY_GOOD_ONES
AS (
SELECT SOME_VALUE
FROM MY_TABLE
WHERE GOOD_ROW = TRUE
)
SELECT *
FROM ONLY_GOOD_ONES
INNER JOIN NUMBERS ON NUMBERS.NUMBER_ID = TO_NUMBER(SOME_VALUE);
Miracle!!!
It worked!
I get my 2 expected records.
So far so good...
However, if I had defined my CTE slightly differently (WHERE clause which filters the same records)
WITH ONLY_GOOD_ONES
AS (
SELECT SOME_VALUE
FROM MY_TABLE
WHERE ROW_ID < 3
)
SELECT *
FROM ONLY_GOOD_ONES;
This CTE returns exactly the same thing as before
But if I try to join, it Fails!
WITH ONLY_GOOD_ONES
AS (
SELECT *
FROM MY_TABLE
WHERE ROW_ID < 3
)
SELECT *
FROM ONLY_GOOD_ONES
INNER JOIN NUMBERS ON NUMBERS.NUMBER_ID = TO_NUMBER(SOME_VALUE);
I get the following error...
SQL Error [100038] [22018]: Numeric value 'ABC' is not recognized
Is there a particular explanation to this second version of the CTE behaving differently???

Answer 1

The actual answer is because snowflake does not follow the SQL standard, and execute SQL in the order given.
They apply transforms to data prior to filtering when there optimizer decides it wants to.
So for your table MY_TABLE when you do
SELECT some_value::NUMBER FROM my_table WHERE row_id IN (1,2);
You will under some cases have the as_number cast happen on all row, and explode on the 'ABC'. Which is violating SQL rules, that WHERE are evaluated before SELECT transforms are done, but Snowflake have known this for years, and it's intentional, as it makes things run faster.
The solution is to understand you have mixed data and therefore assume the code can and will be ran out of order, and thus use the protective versions of the functions like TRY_TO_NUMBER
The kicker is you can write a few nested SELECTs to avoid the problem and then put something like a window funcation around the code and the optimizer jump back into this behavour and you SQL explodes again. Thus the solution is to understand if you have mixed data, and handle it. Oh and complain it's a bug.

Answer 2

This is because you're getting a different execution plan with the different queries.
Here's how the query is executed with the working query:
... and here is how it's executed with the query generating a failure. The error comes from the fact that the join filter is applied directly on the table scan before the ROW_ID < 3 filter is applied, compared to the working query.
You can see these plans under history, clicking the query id and then the 'profile' tab.

Answer 3

It looks like the join filter is applied so early, maybe because of a wrong estimation. When I run the queries on my test database, they completed without any error.
To overcome the issue, you can always "Error-handling Conversion Functions":
SELECT *
FROM MY_TABLE
INNER JOIN NUMBERS ON NUMBERS.NUMBER_ID = TRY_TO_NUMBER(SOME_VALUE)
WHERE ROW_ID < 3;
More information:
https://docs.snowflake.com/en/sql-reference/functions-conversion.html#label-try-conversion-functions