Question

I’m using Oracle 10g. I have a table with a number of fields of varying types. The fields contain observations that have been made by made about a particular thing on a particular date by a particular site.

So:

ItemID, Date, Observation1, Observation2, Observation3...

There are about 40 Observations in each record. The table structure cannot be changed at this point in time.

Unfortunately not all the Observations have been populated (either accidentally or because the site is incapable of making that recording). I need to combine all the records about a particular item into a single record in a query, making it as complete as possible.

A simple way to do this would be something like

SELECT
    ItemID,
    MAX(Date),
    MAX(Observation1),
    MAX(Observation2)
    etc.
FROM
    Table
GROUP BY
    ItemID

But ideally I would like it to pick the most recent observation available, not the max/min value. I could do this by writing sub queries in the form

SELECT
    ItemID,
    ObservationX,
    ROW_NUMBER() OVER (PARTITION BY ItemID ORDER BY Date DESC) ROWNUMBER
FROM
    Table
WHERE
    ObservationX IS NOT NULL

And joining all the ROWNUMBER 1s together for an ItemID but because of the number of fields this would require 40 subqueries.

My question is whether there’s a more concise way of doing this that I’m missing.

Answer 1

Create the table and the sample date

SQL> create table observation(
  2    item_id number,
  3    dt      date,
  4    val1    number,
  5    val2    number );

Table created.

SQL> insert into observation values( 1, date '2011-12-01', 1, null );

1 row created.

SQL> insert into observation values( 1, date '2011-12-02', null, 2 );

1 row created.

SQL> insert into observation values( 1, date '2011-12-03', 3, null );

1 row created.

SQL> insert into observation values( 2, date '2011-12-01', 4, null );

1 row created.

SQL> insert into observation values( 2, date '2011-12-02', 5, 6 );

1 row created.

And then use the KEEP clause on the MAX aggregate function with an ORDER BY that puts the rows with NULL observations at the end. whatever date you use in the ORDER BY needs to be earlier than the earliest real observation in the table.

SQL> ed
Wrote file afiedt.buf

  1  select item_id,
  2         max(val1) keep( dense_rank last
  3                              order by (case when val1 is not null
  4                                             then dt
  5                                             else date '1900-01-01'
  6                                          end) ) val1,
  7         max(val2) keep( dense_rank last
  8                              order by (case when val2 is not null
  9                                             then dt
 10                                             else date '1900-01-01'
 11                                          end) ) val2
 12    from observation
 13*  group by item_id
SQL> /

   ITEM_ID       VAL1       VAL2
---------- ---------- ----------
         1          3          2
         2          5          6

I suspect that there is a more elegant solution to ignore the NULL values than adding the CASE statement to the ORDER BY but the CASE gets the job done.