Indexing Group By

SQL databases use two entirely different group by algorithms. The first one, the hash algorithm, aggregates the input records in a temporary hash table. Once all input records are processed, the hash table is returned as the result. The second algorithm, the sort/group algorithm, first sorts the input data by the grouping key so that the rows of each group follow each other in immediate succession. Afterwards, the database just needs to aggregate them. In general, both algorithms need to materialize an intermediate state, so they are not executed in a pipelined manner. Nevertheless the sort/group algorithm can use an index to avoid the sort operation, thus enabling a pipelined group by.

Consider the following query. It delivers yesterday’s revenue grouped by PRODUCT_ID:

SELECT product_id, sum(eur_value)
  FROM sales
 WHERE sale_date = TRUNC(sysdate) - INTERVAL '1' DAY
 GROUP BY product_id

Knowing the index on SALE_DATE and PRODUCT_ID from the previous section, the sort/group algorithm is more appropriate because an INDEX RANGE SCAN automatically delivers the rows in the required order. That means the database avoids materialization because it does not need an explicit sort operation—the group by is executed in a pipelined manner.

The Oracle database’s execution plan marks a pipelined SORT GROUP BY operation with the NOSORT addendum. The execution plan of other databases does not mention any sort operation at all.

The pipelined group by has the same prerequisites as the pipelined order by, except there are no ASC and DESC modifiers. That means that defining an index with ASC/DESC modifiers should not affect pipelined group by execution. The same is true for NULLS FIRST/LAST. Nevertheless there are databases that cannot properly use an ASC/DESC index for a pipelined group by.

If we extend the query to consider all sales since yesterday, as we did in the example for the pipelined order by, it prevents the pipelined group by for the same reason as before: the INDEX RANGE SCAN does not deliver the rows ordered by the grouping key (compare Figure 6.1).

SELECT product_id, sum(eur_value)
  FROM sales
 WHERE sale_date >= TRUNC(sysdate) - INTERVAL '1' DAY
 GROUP BY product_id

Instead, the Oracle database uses the hash algorithm. The advantage of the hash algorithm is that it only needs to buffer the aggregated result, whereas the sort/group algorithm materializes the complete input set. In other words: the hash algorithm needs less memory.

As with pipelined order by, a fast execution is not the most important aspect of the pipelined group by execution. It is more important that the database executes it in a pipelined manner and delivers the first result before reading the entire input. This is the prerequisite for the advanced optimization methods explained in the next chapter.