Finding empty columns of a table in PostgreSQL

testbed:

create role stack;
create schema authorization stack;
set role stack;

create table my_table as 
select generate_series(0,9) as id, 1 as val1, null::integer as val2;

create table my_table2 as 
select generate_series(0,9) as id, 1 as val1, null::integer as val2, 3 as val3;

function:

create function has_nonnulls(p_schema in text, p_table in text, p_column in text)
                returns boolean language plpgsql as $$
declare 
  b boolean;
begin
  execute 'select exists(select * from '||
          p_table||' where '||p_column||' is not null)' into b;
  return b;
end;$$;

query:

select table_schema, table_name, column_name, 
       has_nonnulls(table_schema, table_name, column_name)
from information_schema.columns
where table_schema='stack';

result:

 table_schema | table_name | column_name | has_nonnulls
--------------+------------+-------------+--------------
 stack        | my_table   | id          | t
 stack        | my_table   | val1        | t
 stack        | my_table   | val2        | f
 stack        | my_table2  | id          | t
 stack        | my_table2  | val1        | t
 stack        | my_table2  | val2        | f
 stack        | my_table2  | val3        | t
(7 rows)

In addition you can get an approximate answer by querying the catalog - if null_frac is zero that indicates no nulls but should be double-checked against 'real' data:

select tablename, attname, null_frac from pg_stats where schemaname='stack';

 tablename | attname | null_frac
-----------+---------+-----------
 my_table  | id      |         0
 my_table  | val1    |         0
 my_table  | val2    |         1
 my_table2 | id      |         0
 my_table2 | val1    |         0
 my_table2 | val2    |         1
 my_table2 | val3    |         0
(7 rows)

In Postgresql, you can get the data directly from the stats:

vacuum analyze; -- if needed

select schemaname, tablename, attname
from pg_stats
where most_common_vals is null
and most_common_freqs is null
and histogram_bounds is null
and correlation is null
and null_frac = 1;

You might get a few false positives, so a recheck is in order after finding the candidates.

I will show you my solution in T-SQL, working for SQL Server 2008. I'm not familiar with PostgreSQL, but I hope that you'll find some guidance in my solution.

-- create test table
IF object_id ('dbo.TestTable') is not null
    DROP table testTable
go
create table testTable (
    id int identity primary key clustered,
    nullColumn varchar(100) NULL,
    notNullColumn varchar(100) not null,
    combinedColumn varchar(100) NULL,
    testTime datetime default getdate()
);
go

-- insert test data:
INSERT INTO testTable(nullColumn, notNullColumn, combinedColumn)
SELECT NULL, 'Test', 'Combination'
from sys.objects
union all
SELECT NULL, 'Test2', NULL
from sys.objects

select *
from testTable

-- FIXED SCRIPT FOR KNOWN TABLE (known structure) - find all completely NULL columns
select sum(datalength(id)) as SumColLength,
    'id' as ColumnName
from dbo.testTable
UNION ALL
select sum(datalength(nullColumn)) as SumColLength,
    'nullColumn' as ColumnName
from dbo.testTable
UNION ALL
select sum(datalength(notNullColumn)) as SumColLength,
    'notNullColumn' as ColumnName
from dbo.testTable
UNION ALL
select sum(datalength(combinedColumn)) as SumColLength,
    'combinedColumn' as ColumnName
from dbo.testTable
UNION ALL
select sum(datalength(testTime)) as SumColLength,
    'testTime' as ColumnName
from dbo.testTable

-- DYNAMIC SCRIPT (unknown structure) - find all completely NULL columns
declare @sql varchar(max) = '', @tableName sysname = 'testTable';

SELECT @sql +=
        'select sum(datalength(' + c.COLUMN_NAME + ')) as SumColLength,
    ''' + c.COLUMN_NAME + ''' as ColumnName
from ' + c.TABLE_SCHEMA + '.' + c.TABLE_NAME --as StatementToExecute
+ '
UNION ALL
'
FROM INFORMATION_SCHEMA.COLUMNS c
WHERE c.TABLE_NAME = @tableName;

SET @sql = left(@sql, len(@sql)-11)
print @sql;
exec (@sql);

What I did, in short, was to create a test table with 5 columns, ID and testTime being generated by identity and getdate() function, while the 3 varchar columns being the ones of interest. One will have only NULL values, one will not have any NULLs, the other will be a combined column. The final result of the script will be that the script will report the column nullColumn as having all rows NULL.

The idea was to calculate the function DATALENGTH for each column (calculates the number of bytes for a given expression). So I calculated the DATALENGTH value for each row of each column and made a SUM per column. If the SUM per column is NULL, then the complete column has NULL rows, otherwise there is some data inside.

Now you have to find the translation for PostgreSQL and hopefully a colleague will be able to help you with that. Or maybe there is a nice system view that will show how dumb I am for reinventing the wheel :-).

You need to query the information catalog for such information:

SELECT column_name FROM information_schema.columns WHERE table_name='your_table'

gives you the matching tables for your columns.

I don't have a postgres installation currently at hand but the rest should be simple

   loop over the results of the above query and foreach result
        send a COUNT(*) to the table
        if the count is null, give back the column,
                 else ignore it
   end foreach