SELECT TOP in Oracle

Top-N queries are queries that limit the result to a specific number of rows. These are often queries for the most recent or the best entries of a result set. For efficient execution, the ranking must be done with a pipelined order by.

The simplest way to fetch only the first rows of a query is fetching the required rows and then closing the statement. Unfortunately, the optimizer cannot foresee that when preparing the execution plan. To select the best execution plan, the optimizer has to know if the application will ultimately fetch all rows. In that case, a full table scan with explicit sort operation might perform best, although a pipelined order by could be better when fetching only ten rowseven if the database has to fetch each row individually. That means that the optimizer has to know if you are going to abort the statement before fetching all rows so it can select the best execution plan.

The SQL standard excluded this requirement for a long time. The corresponding extension (fetch first) was finally introduced with SQL:2008 and is currently available in IBM DB2, PostgreSQL, SQL Server 2012 and Oracle 12c. On the one hand, this is because the feature is a non-core extension, and on the other hand its because each database has been offering its own proprietary solution for many years.

The following examples show the use of these well-known extensions by querying the ten most recent sales. The basis is always the same: fetching all sales, beginning with the most recent one. The respective top-N syntax just aborts the execution after fetching ten rows.

All of the above shown SQL queries are special because the databases recognize them as top-N queries.

If the optimizer is aware of the fact that we only need ten rows, it will prefer to use a pipelined order by if applicable:

The Oracle execution plan indicates the planned termination with the COUNT STOPKEY operation. That means the database recognized the top-N syntax.

If there is no suitable index on SALE_DATE for a pipelined order by, the database must read and sort the entire table. The first row is only delivered after reading the last row from the table.

This execution plan has no pipelined order by and is almost as slow as aborting the execution from the client side. Using the top-N syntax is still better because the database does not need to materialize the full result but only the ten most recent rows. This requires considerably less memory. The Oracle execution plan indicates this optimization with the STOPKEY modifier on the SORT ORDER BY operation.

The advantages of a pipelined top-N query include not only immediate performance gains but also improved scalability. Without using pipelined execution, the response time of this top-N query grows with the table size. The response time using a pipelined execution, however, only grows with the number of selected rows. In other words, the response time of a pipelined top-N query is always the same; this is almost independent of the table size. Only when the B-tree depth grows does the query become a little bit slower.

Figure7.1 shows the scalability for both variants over a growing volume of data. The linear response time growth for an execution without a pipelined order by is clearly visible. The response time for the pipelined execution remains constant.

Figure7.1 Scalability of Top-N Queries

Although the response time of a pipelined top-N query does not depend on the table size, it still grows with the number of selected rows. The response time will therefore double when selecting twice as many rows. This is particularly significant for paging queries that load additional results because these queries often start at the first entry again; they will read the rows already shown on the previous page and discard them before finally reaching the results for the second page. Nevertheless, there is a solution for this problem as well as we will see in the next section.