sql: use separate lookahead of WITH for BUCKET_COUNT

[chengxiong-ruan]

Fixes #75604

In #75231 we use the WITH_LA to resolve conflicts on the
WITH BUCKET_COUNT clause. I suspect it somehow made the
random sql generator generate sqls like this:

IMPORT INTO FAMILY . FAMILY . CHAR ( ident , ident ) PUBLICATION DATA ( PLACEHOLDER ) WITH OPTIONS ( BUCKET_COUNT = PLACEHOLDER );

which is formatted into:

IMPORT INTO "family"."family"."char"(ident, ident) PUBLICATION DATA ('placeholder') WITH bucket_count = 'placeholder'

this fails the parser for sure, since it's not a legit statement in crdb.
This PR creates a new lookahead token to prevent the sql generator
from generating IMPORT statement with BUCKET_COUNT in it.

Release note: None

[cockroach-teamcity]

This change is 

[chengxiong-ruan]

To be clear that I ran a roachprod stress to repro but I was just not lucky enough to make it fail on the same case. So I wasn't able to prove my theory with a it failed before but fixed after verification :(

Reviewable status: complete! 0 of 0 LGTMs obtained (waiting on @ajwerner and @rafiss)

[rafiss]

thanks for fixing this! can you add a regression test case in pkg/sql/parser/testdata?

[rafiss]

i.e. apply this patch

diff --git a/pkg/sql/parser/testdata/import_export b/pkg/sql/parser/testdata/import_export
index e7f367c4c2..1223d1c033 100644
--- a/pkg/sql/parser/testdata/import_export
+++ b/pkg/sql/parser/testdata/import_export
@@ -85,3 +85,7 @@ EXPORT INTO CSV 's3://my/path/%part%.csv' WITH delimiter = '|' FROM SELECT a, su
 EXPORT INTO CSV ('s3://my/path/%part%.csv') WITH delimiter = ('|') FROM SELECT (a), ((sum)((b))) FROM c WHERE ((d) = (1)) ORDER BY ((sum)((b))) DESC LIMIT (10) -- fully parenthesized
 EXPORT INTO CSV '_' WITH delimiter = '_' FROM SELECT a, sum(b) FROM c WHERE d = _ ORDER BY sum(b) DESC LIMIT _ -- literals removed
 EXPORT INTO CSV 's3://my/path/%part%.csv' WITH _ = '|' FROM SELECT _, sum(_) FROM _ WHERE _ = 1 ORDER BY sum(_) DESC LIMIT 10 -- identifiers removed
+
+parse
+IMPORT INTO FAMILY . FAMILY . CHAR ( ident , ident ) PUBLICATION DATA ( PLACEHOLDER ) WITH OPTIONS ( BUCKET_COUNT = PLACEHOLDER );
+----

then run make test PKG=./pkg/sql/parser TESTS='TestParseDatadriven' TESTFLAGS='-rewrite'

[chengxiong-ruan]

@rafiss thanks for the hint, but the issue here is the random sql generator generated unexpected sql statement. so the statement would fail anyway. I played with the test you mentioned above and it just failed with parse_test.go:47: at or near "with": syntax error.

[rafiss]

Ah I see, then I'm not sure this fix will change that -- the opt_with_options rule in sql.y could still lead to an option with a name of an unreserved_keyword to be used. Maybe we need to update the random syntax generator itself.

[chengxiong-ruan]

oh yeah, agreed... I just suspect my previous change cause it so hope this change would solve it.
Now looks like it's probably just the timing that we were lucky enough and never hit into the failure before, but only after my change merged. Closing this pr for now. Do you know who might be the best person to ask for advice how the random generator can be fixed?

[rafiss]

Yeah looks like unlucky timing with your PR! Since it wasn't caused by that, feel free to reassign the issue back to me.

But if you would still like to keep working on it, that would be very appreciated :)

Do you know who might be the best person to ask for advice how the random generator can be fixed?

No one's worked on it for a while unfortunately. Asking in #sql-experience might be the best way.

[chengxiong-ruan]

I probably don't have enough bandwidth at the moment :( Assigning back to you for now. Thanks!

Reviewable status: complete! 0 of 0 LGTMs obtained (waiting on @ajwerner)