On Saturday, March 17, 2012 11:04:30 PM Tom Lane wrote: > I've found a couple more issues in the CTAS patch: > > 1. Previous versions delivered a "SELECT n" command tag for either > spelling of the command: > > regression=# select * into t1 from int8_tbl; > SELECT 6 > regression=# create table t2 as select * from int8_tbl; > SELECT 6 > > With the patch I get > > regression=# select * into t1 from int8_tbl; > SELECT 0 0 hm. Stupid me.
> regression=# create table t2 as select * from int8_tbl; > CREATE TABLE AS > > The first of these is particularly unfortunate since it's outright lying > as to the number of rows processed. > I'm not sure what we should do instead. We have gotten push-back before > anytime we changed the command tag for an existing command (and in fact > it seems that we intentionally added the rowcount display in 9.0, which > means there are people out there who care about that functionality). > On the other hand, the traditional output for CREATE TABLE AS doesn't > seem to satisfy the principle of least astonishment. A third > consideration is that if we are pushing CREATE TABLE AS as the preferred > spelling, people will probably complain if it omits functionality that > SELECT INTO provides; so I'm not sure that "SELECT n" in one case and > "CREATE TABLE AS" in the other would be a good idea either. Any > opinions what to do here? I would prefer both returning CREATE TABLE AS since thats what actually happens. We already document that SELECT INTO is kinda deprecated: "The PostgreSQL usage of SELECT INTO to represent table creation is historical. It is best to use CREATE TABLE AS for this purpose in new code." I have seen code that uses the command code for selecting the, app level, logging. Its kinda hard to do that if a CREATE TABLE AS/SELECT INTO returns SELECT. Does CTAS ommit any functionality currently? I don't see any reason not to support stuff there thats supported in SELECT INTO. > 2. Historically, CREATE RULE has allowed a rule action to be SELECT INTO > (though not CREATE TABLE AS). Currently the patch is throwing an error > for that. This seems like something that might not be worth fixing, > though. It's fairly hard to conceive of a use-case for such a rule, > since it would work only once before starting to throw "table already > exists" errors. How much do we care about preserving backward > compatibility here? I vote for not supporting that anymore. That being possible looks more like an implementation detail to me. Thanks for looking at this! Andres -- Sent via pgsql-hackers mailing list (email@example.com) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers