I am wondering if there is a way to copy a csv file into a temp table where the number of columns is unknown in the csv file. The DB software I'm using is PgAdmin III. I figured out that if I do know the number of columns then I can create a temp table with that amount of columns and then copy the csv file in like so:

      col1 VARCHAR(80),
      col2 VARCHAR(80),
      coln VARCHAR(80)

COPY temp FROM 'C:/Users/postgres/Boost.txt' CSV HEADER DELIMITER E'    '

However if I try to just simply copy the csv file to a temp table with no columns in the temp table, Postgresql (version 8.4) complains that I'm working with a table that has less columns than in the csv file. I've been researching and can't seem to find anything in the Postgresql docs about this. Does anyone know if copying a csv file into a temp table with an arbitrary number of columns decided at run time is possible in Postgresql? Once the temp table is loaded with the csv file I plan on doing some comparisons to other tables with the temp table before it is destroyed. Also the first row in the csv file contains headers.

  • 1
    You'll want to use an ETL tool or script. Take a look at CloverETL, Talend Studio, Pentaho Kettle. There are probably also a bunch of simple "CSV loader" scripts out there for your preferred scripting language. – Craig Ringer Jul 1 '15 at 1:29


  • The 1st row of the CSV file has column names of defined format.
  • The PROGRAM clause of COPY and GET DIAGNOSTICS after COPY require Postgres 9.3+.
  • format() requires Postgres 9.1+
  • This works with pure standard Postgres - except for the head command that the shell is expected to provide. For Windows versions consider:

Full automation

This function copies any table structure completely dynamically:

CREATE OR REPLACE FUNCTION f_dynamic_copy(_file    text
                                        , _tbl     text = 'tmp1'
                                        , _delim   varchar(1) = E'\t'
                                        , _nodelim varchar(1) = '')
   row_ct int;
   -- create staging table for 1st row as single text column 

   -- fetch 1st row
   EXECUTE format($$COPY tmp0 FROM PROGRAM 'head -n1 %I' WITH (DELIMITER %L)$$
                , _file, _delim);

   -- create actual temp table with all columns text
      SELECT format('CREATE TEMP TABLE %I(', _tbl)
          || string_agg(quote_ident(col) || ' text', ',')
          || ')'
      FROM  (SELECT cols FROM tmp0 LIMIT 1) t
           , unnest(string_to_array(t.cols, E'\t')) col

   -- Import data
                , _tbl, _file, _delim);

   RETURN format('%I with %s rows.', _tbl, row_ct);
$func$  LANGUAGE plpgsql;


SELECT f_dynamic_copy('/path/to/file.csv');


SELECT f_dynamic_copy('/path/to/file2.csv', 'tmp_file2');


tmp_file2 with 10 rows.

Before the main COPY, run a preliminary COPY ... TO tmp0 to fetch the first row with column names, which are expected to be unquoted, case-sensitive strings like COPY ... TO ... (FORMAT csv, HEADER) would export them.

The structure of the actual target table is derived from it, all columns with data type text. The default name of the resulting table is tmp1 - unless you provide a table name as 2nd parameter to the function.

Then COPY is executed. The default delimiter is a tab character - or provide your delimiter as 3rd parameter to the function.

Use any single-byte character for the non-delimiter _nodelim which does not appear in the first line of your CSV file. I am arbitrarily picking the control character "Delete" (ASCII 127), assuming that won't pop up - or provide your non-delimiter as 4th parameter to the function.

The function returns table name and number of imported rows.
Remember, a temporary table dies with the end of the session.

Per documentation:

Executing a command with PROGRAM might be restricted by the operating system's access control mechanisms, such as SELinux.

Related answer on SO:

Postgres 8.4

That version is too old, I am not going to back-port that far.

  • GET DIAGNOSTICS is an optional feature. You can just leave it away or replace it with a full count on the table

  • A primitive (expensive) alternative for the PROGRAM clause of COPY in pg 9.3 would be to import the complete table instead:

    EXECUTE format($$COPY tmp0 FROM %L WITH (DELIMITER %L)$$, _file, _delim);

Or you prepare a second input file, or you can make it work by piping from the shell: COPY tablename FROM STDIN is available in pg 8.4.

  • Well I don't have the option of upgrading at this time. So the line of code COPY tablename FROM STDIN is what I would use to replace all instances of EXECUTE format... or do I need it for just fetching the first row as the column names and the rest of the rows as data? – Bmoe Jul 1 '15 at 16:40
  • @John: Only for the first instance, the point of which is to fetch the first row only. You could even COPY the complete table another time, it just would be twice the cost. The second instance is a plain COPY (dynamically using input parameters). I also added info on format(). You should urgently consider upgrading to a current version. – Erwin Brandstetter Jul 1 '15 at 16:48

This program handles creating PostgreSQL tables based on most any CSV or JSON with one line:


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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