Note: before this fix, Default value expression would fall to UnresolvedAttribute above. It would then think the default value refers to the conflicting column name and fail.

case u @ UnresolvedAttribute(nameParts) =>
          val result = withPosition(u) {
            resolveColumnByName(nameParts)
              .orElse(LiteralFunctionResolution.resolve(nameParts))
              .map {
                // We trim unnecessary alias here. Note that, we cannot trim the alias at top-level,
                // as we should resolve `UnresolvedAttribute` to a named expression. The caller side
                // can trim the top-level alias if it's safe to do so. Since we will call
                // CleanupAliases later in Analyzer, trim non top-level unnecessary alias is safe.
                case Alias(child, _) if !isTopLevel => child
                case other => other
              }
              .getOrElse(u)
          }
          logDebug(s"Resolving $u to $result")
          result

I just copied this from the other code, we dont need it right?

I assume this works because we resolve expressions top to bottom, hence we see DefaultValueExpression before we see the unresolved attribute?

I am a little bit worried that this doesn't fix the root problem that ResolvedIdentifier output in CREATE actually is used as candidates for resolving default values. Let me think.

Actually, maybe it does solve the root problem. Am I right we will not recurse into DefaultValueExpression child so we effectively ensure that default value resolution doesn't have access to attributes?

yes, this is matched before children

Do we have tests when a default value references other data columns? It is illegal, just to want to make sure we throw a good error message in this case.

col1 INT,
col2 INT DEFAULT col1

yes , do you mean like "test default value should not refer to real column", later in the file?