I did some investigating and it looks like if we modify _spanner_temp_table_keyword_args to return {"prefixes": ["TEMPORARY"]} then we should be able to differentiate between real and temporary tables in post_create_table by checking for the TEMPORARY prefix. This means we may still be able to raise an Unimplemented error.

From local testing, this seems to prevent creating a temporary table via the TEMPORARY prefix when declaring the table. However, the test suite started having failures which I'm guessing are from tests using temporary tables. This may mean more overrides of define_tables would be necessary.

@skuruppu WDYT?

Yeah I think ideally we would prevent anyone from attempting to create temporary tables since they won't behave the way a user might expect when using real tables. So returning UNIMPLEMENTED is ideal if it's not too complicated to implement.

As for tests, how many tests will fail if we go down this path? Would it mean that we have to skip a large portion of the test suite or is it just a few tests? If it's a few, can we override the tests to create real tables?

@skuruppu, I'd expect about 60% of tests failing. It's about 120 test cases.
Other don't use tables, just work with the dialect itself.
Running now the whole suite to get the real number.

UPDATE: about 40% of tests failed. Looks like our previous test overrides are helping us here. Looking at how the failed tests are organized, I think, we can override them fast enough. Doing it.

@skuruppu, @larkee, yes, this idea with catching prefixes in post_create_table works fine 👍 Only one test case had to be updated (reflection - a very big bunch of tests, which gives 40% of the whole suite). Pushing...