Feat(teradata): add support for the SAMPLE clause

[georgesittas]

Motivated by #2161

[Ashay4u]

Pardon my limited understanding here, but where is tablesample_sql from teradata.py called from? I did not reach that function in my copy of the code

[georgesittas]

@Ashay4u it's called from the base generator's sql method when calling transpile on a query that contains the SAMPLE clause. What do you mean you didn't reach that function?

[Ashay4u]

I added a breakpoint on def tablesample_sql and its return statement in terdata.py file and the code did not stop there during debug. It does not arrive at the breakpoint for tablesample_sql in generator.py during transpile too. Am i missing a change somewhere
Also the output that I see after running the code is as below. Notice there are no values in the sample clause

SELECT
  *
FROM Employee Sample ()

[georgesittas]

Can you share your example so I can try to reproduce this? What I've tried so far reaches that code.

[Ashay4u]

Code below

sql="""
SELECT * FROM 
	Employee 
	Sample 0.33, .25, .1;
"""

try:
    new_query = sqlglot.transpile(sql, read='teradata', write='databricks', pretty=True)[0]
    print(new_query)
except sqlglot.errors.ParseError as error:
    print(traceback.format_exc())
    print(error.errors)

My code does not stop at tablesample_sql in generator.py too during transpile. Am i missing a change somewhere

[georgesittas]

Yeah, it doesn't reach that method because you're transpiling to databricks. That doesn't work currently, I only added support for parsing/generating in Teradata, because the transpilation is out of scope for now.

See related comment: #2161 (comment).

[Ashay4u]

Thanks for clarifying. Where should I make changes to the code to try and fix this? I will try to get this to some acceptable fix

[georgesittas]

Ideally you need to parse this into a canonical representation, so that:

1. You don't lose information when transpiling to Teradata, i.e. you need to preserve SAMPLE 0.33, 0.25, 0.1.
2. You can generate valid SQL when targeting other dialects from that canonical representation.

From what I saw, in order to implement (2), you need to check what kind of numbers you have in the SAMPLE. In your example, 0.33 would be converted to 33 PERCENT, whereas SAMPLE 4 would probably be converted into 4 ROWS.

I see you mentioned ... TABLESAMPLE 33 PERCENT as the expected Databricks output. Why is that? Aren't you dropping 0.25 and 0.1 in that case, or is that an acceptable information loss?

[Ashay4u]

For my use case, its an acceptable loss. Is there an easy fix for it?

[georgesittas]

The fix is not trivial, at least if you're not experienced with SQLGlot's codebase. One idea is to override _parse_table_sample in Teradata and:

• Call super()._parse_table_sample
• If that returns a TableSample, you can check the first arg of expressions and depending on what kind of number it is you can also set another arg like percent, rows etc.

See my edit in the above reply too.

[Ashay4u]

Thanks. Will give it a try. How do I familiarize myself with the codebase details. Is there documentation I can refer or just debug through the code to understand? Again, thanks a lot for being so prompt and helpful

[georgesittas]

Sure thing, glad you want to contribute. You can also join our slack btw if you want to chat.

How do I familiarize myself with the codebase details. Is there documentation I can refer or just debug through the code to understand?

Yeah I'd suggest debugging through the code and looking into past issues & their fixes to understand the work flow. I think the above outline should cover your use case, so try to work around that.