fix: remove pre-caching of remote function results

[shobsi]

This will save redundant remote function execution as reported in b/370088754. The trade-off is that any remote function integration issues will be caughts only at a later point through a usage triggered sql execution. Why the caching is not working as indended in the reported use case in the bug? as per TrevorBergeron@ "the cached execution is useless when it is implicitly joined back to the base dataframe"

Thank you for opening a Pull Request! Before submitting your PR, there are a few things you can do to make sure it goes smoothly:

• Make sure to open an issue as a bug/issue before writing your code! That way we can discuss the change, evaluate designs, and agree on the general idea
• Ensure the tests and linter pass
• Code coverage does not decrease (if any source code was changed)
• Appropriate docs were updated (if necessary)

Fixes internal issue 370088754 🦕

[GarrettWu]

So after the change, when does the execution happens?

df["col2"] = df["col1"].apply(remote_function)

or

df.to_gbq()

if at the later stage, if the user calls exec APIs (such as repr) multiple times, will the function be executed multiple times?

[GarrettWu]

Is there a way that we can add test to ensure that with multiple operations, there won't be redundant function calls?

questions

[shobsi]

So after the change, when does the execution happens?

df["col2"] = df["col1"].apply(remote_function)

or

df.to_gbq()

if at the later stage, if the user calls exec APIs (such as repr) multiple times, will the function be executed multiple times?

It would be the latter. After this change the behavior would be same as the standard bigframes behavior - if the result is cached it won't be executed, else it would be.

[shobsi]

Is there a way that we can add test to ensure that with multiple operations, there won't be redundant function calls?

It's tied to caching. Chatted with @TrevorBergeron that caching is not a guarantee at this point but more like an internal optimization. Asserting how many times a function was called is a bit of a stretch, just like asserting how many number of sql jobs were run in an operation, so I'd skip adding a test. Let me know if you have s strong opinion otherwise.