I want to pass a table name as a parameter in a Postgres function. I tried this code:

CREATE OR REPLACE FUNCTION some_f(param character varying) RETURNS integer 
AS $$
    BEGIN
    IF EXISTS (select * from quote_ident($1) where quote_ident($1).id=1) THEN
     return 1;
    END IF;
    return 0;
    END;
$$ LANGUAGE plpgsql;

select some_f('table_name');

And I got this:

ERROR:  syntax error at or near "."
LINE 4: ...elect * from quote_ident($1) where quote_ident($1).id=1)...
                                                             ^

********** Error **********

ERROR: syntax error at or near "."

And here is the error I got when changed to this select * from quote_ident($1) tab where tab.id=1:

ERROR:  column tab.id does not exist
LINE 1: ...T EXISTS (select * from quote_ident($1) tab where tab.id...

Probably, quote_ident($1) works, because without the where quote_ident($1).id=1 part I get 1, which means something is selected. Why may the first quote_ident($1) work and the second one not at the same time? And how could this be solved?

upvote
  flag
I know this question is kind of old, but I found it while searching for the answer to another issue. Couldn't your function just query the informational_schema? I mean, that's kind of what it's for in a way - to let you query and see what objects exist in the database. Just an idea. – David S
upvote
  flag
@DavidS Thanks for a comment, I'll try that. – John Doe

6 Answers 11

The first doesn't actually "work" in the sense that you mean, it works only in so far as it does not generate an error.

Try SELECT * FROM quote_ident('table_that_does_not_exist');, and you will see why your function returns 1: the select is returning a table with one column (named quote_ident) with one row (the variable $1 or in this particular case table_that_does_not_exist).

What you want to do will require dynamic SQL, which is actually the place that the quote_* functions are meant to be used.

upvote
  flag
Thanks a lot, Matt, table_that_does_not_exist gave the same result, you're right. – John Doe

Inside plpgsql code, The EXECUTE statement must be used for queries in which table names or columns come from variables. Also the IF EXISTS (<query>) construct is not allowed when query is dynamically generated.

Here's your function with both problems fixed:

CREATE OR REPLACE FUNCTION some_f(param character varying) RETURNS integer 
AS $$
DECLARE
 v int;
BEGIN
      EXECUTE 'select 1 FROM ' || quote_ident(param) || ' WHERE '
            || quote_ident(param) || '.id = 1' INTO v;
      IF v THEN return 1; ELSE return 0; END IF;
END;
$$ LANGUAGE plpgsql;
upvote
  flag
Thank you, I was making just the same a couple of minutes ago when read your answer. The only difference is I had to remove quote_ident() because it added extra quotes, which surprised me a little, well, 'cause it is used in most examples. – John Doe
upvote
  flag
Those extra quotes will be needed if/when the table name contains characters outside [a-z], or if/when it clashes with a reserved identifier (example: "group" as a table name) – Daniel Vérité
upvote
  flag
And, by the way, could you please provide a link that would prove that IF EXISTS <query> construct doesn't exist? I'm pretty sure I saw something like that as a working code sample. – John Doe
upvote
  flag
@JohnDoe: IF EXISTS (<query>) THEN ... is a perfectly valid construct in plpgsql. Just not with dynamic SQL for <query>. I use it a lot. Also, this function can be improved quite a bit. I posted an answer. – Erwin Brandstetter
1 upvote
  flag
Sorry, you're right about if exists(<query>), it's valid in the general case. Just checked and modified the answer accordingly. – Daniel Vérité
upvote
  flag
@john 'SELECT 1' inside an EXISTS clause does not create a 1-column table. It's time to stop correcting professionals with rudimentary and unlearned reasoning. – ErikE
upvote
  flag
@ErikE Thanks for mentioning that (though rather emotionally). – John Doe
up vote 72 down vote accepted

This can be further simplified and improved:

CREATE OR REPLACE FUNCTION some_f(_tbl regclass, OUT result integer) AS
$func$
BEGIN
EXECUTE format('SELECT (EXISTS (SELECT 1 FROM %s WHERE id = 1))::int', _tbl)
INTO result;
END
$func$ LANGUAGE plpgsql;

Call (example with schema-qualified name - see below):

SELECT some_f('myschema.mytable');  -- would fail with quote_ident()

Or:

SELECT some_f('"my very uncommon table name"')

Major points

  • Use an OUT parameter to simplify the function. You can directly select the result of the dynamic SQL into it and be done. No need for additional variables and code.

  • EXISTS does exactly what you want, I kept it in my query. There are various ways to do this.

  • You seem to want an integer back, so I cast the boolean result from EXISTS() to integer, which yields exactly what you had. I would return boolean instead.

  • I use the object identifier type regclass as input type for _tbl. That does everything quote_ident(_tbl) or format('%I', _tbl) would do, but better, because:

    • .. it prevents SQL injection just as well.

    • .. it fails immediately and more gracefully if the table name is invalid / does not exist / is invisible to the current user.

    • .. it works with schema-qualified table names, where a plain quote_ident(_tbl) or format(%I) would fail because they cannot resolve the ambiguity.

  • I still use format(), because it simplifies the syntax (and to demonstrate how it's used), but with %s instead of %I. For the simple example we could as well just concatenate:

    EXECUTE 'SELECT (EXISTS (SELECT 1 FROM ' || _tbl || ' WHERE id = 1))::int'
    
  • No need to table-qualify the id column.

Tested with PostgreSQL 9.1. format() requires at least that version.

This is why you always escape user input for dynamic SQL properly:
SQL Fiddle demo for SQL injection

upvote
  flag
Looks neat, thanks. – John Doe
1 upvote
  flag
Can you execute an analyze command? – suhprano
2 upvote
  flag
@suhprano: Sure. Try it: DO $$BEGIN EXECUTE 'ANALYZE mytbl'; END$$; – Erwin Brandstetter
upvote
  flag
why %s and not %L? – Lotus
upvote
  flag
@Lotus: The explanation is in the answer. regclass values are escaped automatically when output as text. %L would be wrong in this case. – Erwin Brandstetter

Don't do this.

That's the answer. It's a terrible anti-pattern. What purpose does it serve? If the client knows the table it wants data from, then SELECT FROM ThatTable! If you've designed your database in a way that this is required, you've probably designed it wrong. If your data access layer needs to know if a value exists in a table, it is trivially easy to do the dynamic SQL part in that code. Pushing it into the database is not good.

I have an idea: let's install a device inside elevators where you can type in the number of the floor you want. Then when you press "Go", it moves a mechanical hand over to the correct button for the desired floor and presses it for you. Revolutionary!

Apparently my answer was too short on explanation so I am repairing this defect with more detail.

I had no intention of mockery. My silly elevator example was the very best device I could imagine for succinctly pointing out the flaws of the technique suggested in the question. That technique adds a completely useless layer of indirection, and needlessly moves table name choice from a caller space using a robust and well-understood DSL (SQL) into a hybrid using obscure/bizarre server-side SQL code.

Such responsibility splitting through movement of query construction logic into dynamic SQL makes the code harder to understand. It destroys a perfectly reasonable convention (how a SQL query chooses what to select) in the name of custom code fraught with potential for error.

  • Dynamic SQL offers the possibility of SQL injection that is hard to recognize in the front end code or the back end code singly (one must inspect them together to see this).

  • Stored procedures and functions can access resources that the SP/function owner has rights to but the caller doesn't. As far as I understand, when you use code that produces dynamic SQL and runs it, the database executes the dynamic SQL under the rights of the caller. This means you either won't be able to use privileged objects at all, or you have to open them up to all clients, increasing the surface area of potential attack to privileged data. Setting the SP/function at creation time to always run as a particular user (in SQL Server, EXECUTE AS) may solve that problem, but makes things more complicated. This exacerbates the risk of SQL injection mentioned in the previous point, by making the dynamic SQL a very enticing attack vector.

  • When a developer must understand what the application code is doing in order to modify it or fix a bug, he'll find it very difficult to get the exact SQL query being executed. SQL profiler can be used, but this takes special privileges and can have negative performance effects on production systems. The executed query can be logged by the SP but this increases complexity for no reason (maintaining new tables, purging old data, etc.) and is totally non-obvious. In fact, some applications are architected such that the developer does not have database credentials, so it becomes almost impossible for him to actually see the query being submitted.

  • When an error occurs, such as when you try to select a table that doesn't exist, you'll get a message along the lines of "invalid object name" from the database. That will happen exactly the same whether you're composing the SQL in the back end or the database, but the difference is, some poor developer who's trying to troubleshoot the system has to spelunk one level deeper into yet another cave below the one where the problem actually exists, to dig into the wonder-procedure that Does It All and try to figure out what the problem is. Logs won't show "Error in GetWidget", it will show "Error in OneProcedureToRuleThemAllRunner". This abstraction will just make your system worse.

Here's a far better example in pseudo-C# of switching table names based on a parameter:

string sql = string.Format("SELECT * FROM {0};", escapeSqlIdentifier(tableName));
results = connection.Execute(sql);

Every flaw I have mentioned with the other technique is completely absent from this example.

There's just no purpose, no benefit, no possible improvement in submitting a table name to a stored procedure.

2 upvote
  flag
I don't completely agree with that. Say, you press this "Go" button and then some mechanism checks, if the floor exists. Functions may be used in triggers, which in turn can check some conditions. This desicion may not be the most beautiful, but if the system is big enough already and you need to make some corrections in its logic, well, this choice is not so dramatic, I suppose. – John Doe
upvote
  flag
But consider that the action of trying to press a button that doesn't exist will simply generate an exception no matter how you handle it. You can't actually push a nonexistent button, so there is no benefit to adding, on top of button-pushing, a layer to check for nonexistent numbers, since such number entry didn't exist before you created said layer! Abstraction is in my opinion the most powerful tool in programming. However, adding a layer that merely poorly duplicates an existing abstraction is wrong. The database itself is already an abstraction layer that maps names to data sets. – ErikE
1 upvote
  flag
Spot on. The whole point of SQL is to express the set of data that you want extracted. The only thing that this function does is to encapsulate a "canned" SQL statement. Given the fact that the identifier is also hard coded the whole thing has a bad smell to it. – Nick Hristov
upvote
  flag
What about a package to guess the database structure, tablenames, etc, and at some moment it also wants to known the number of rows in the tables it finds. – arivero
upvote
  flag
@arivero "a package"? What are we talking about here? Can't the package construct its query using the known table name? You're not proposing anything different than what I already addressed. – ErikE
upvote
  flag
I have a database provided by a 3rd party application. I want to find all the tables that have been updated since a timestamp on a record in that database. I think I can do that with an EXECUTE, but obviously that's not the correct way. What is the correct way? – AixNPanes
upvote
  flag
Please ask a new question. Comments aren't a good way to get answers. – ErikE
upvote
  flag
This is the same logic that was used to scare everybody away from gotos in C code. It's generally good advice, but it's not always a bad thing. – three-cups
upvote
  flag
@three Until someone is in the mastery phase (see the Dreyfus model of skill acquisition) of a skill, he should simply absolutely obey rules like "do NOT pass table names into a procedure to be used in dynamic SQL". Even hinting that it's not always bad is is itself bad advice. Knowing this, the beginner will be tempted to use it! That's BAD. Only masters of a topic should be breaking the rules, as they're the only ones with the experience to know in any particular case whether such rule-breaking actually makes sense. – ErikE
upvote
  flag
@ErikE You're right, you should know what you're doing before using a non-standard approach. I might focus more energy into explaining why it's a bad idea instead of mocking technique in question. – three-cups
1 upvote
  flag
@three-cups I did update with a lot more detail on why it's a bad idea. – ErikE
upvote
  flag
"At least in SQL Server, dynamic SQL executes with the privileges of the caller, not the running code." - what do you mean with that? Typically any statement you run, runs with the privileges of the caller. And besides: in Postgres (and SQL Server) you can define under which "privileges" a function is executed: either the owner/creator of the function or the current user (=caller) – a_horse_with_no_name
upvote
  flag
@a_horse_with_no_name Updated with a much better description for the privilege concerns I was thinking of. – ErikE
upvote
  flag
Well, in Postgres dynamic SQL runs with the privileges that were defined for the function. You can't break out of that, just because of dynamic SQL. But I still don't get your concerns related to that. Executing the code with the privileges of the caller is what the caller will typically expect. And if you create a function that runs with elevated privileges (which is what "security definer" usually is used for) then you shouldn't be allowing the caller to pass the SQL to be executed. – a_horse_with_no_name
upvote
  flag
Aaaaand the accepted answer does just that: accepts a "table name" which is dropped right in the middle of an SQL statement and executed, and may very well not be a table name at all. Such as Little Bobby Tables showed us all. – ErikE

I have 9.4 version of PostgreSQL and I always use this code:

CREATE FUNCTION add_new_table(text) RETURNS void AS
$BODY$
begin
    execute
        'CREATE TABLE ' || $1 || '(
        item_1      type,
        item_2      type
        )';
end;
$BODY$
LANGUAGE plpgsql

And then:

SELECT add_new_table('my_table_name');

It works good for me.

Attention! Above example is one of those which shows "How do not if we want to keep safety during querying the database" :P

upvote
  flag
Creating a new table is different from operating with the name of an existing table. Either way, you should escape text parameters executed as code or you are open to SQL injection. – Erwin Brandstetter
upvote
  flag
Oh, yeah, my mistake. The topic misled me and in addition I didn't read it to the end. Normally in my case. :P Why code with a text parameter is exposed to injection? – dm3
1 upvote
  flag
upvote
  flag
Oops, it's really dangerous. Thank you for the answer! – dm3

If the question was to test if the table is empty or not (id=1), here is a simplified version of Erwin's stored proc :

CREATE OR REPLACE FUNCTION isEmpty(tableName text, OUT zeroIfEmpty integer) AS
$func$
BEGIN
EXECUTE format('SELECT COALESCE ((SELECT 1 FROM %s LIMIT 1),0)', tableName)
INTO zeroIfEmpty;
END
$func$ LANGUAGE plpgsql;

Not the answer you're looking for? Browse other questions tagged or ask your own question.