1 of 1

Semi-join Materialization Strategy

Semi-join Materialization is a special kind of subquery materialization used for Semi-join subqueries. It actually includes two strategies:

Materialization/lookup
Materialization/scan

The idea

Consider a query that finds countries in Europe which have big cities:

The subquery is uncorrelated, that is, we can run it independently of the upper query. The idea of semi-join materialization is to do just that, and fill a temporary table with possible values of the City.country field of big cities, and then do a join with countries in Europe:

The join can be done in two directions:

From the materialized table to countries in Europe
From countries in Europe to the materialized table

The first way involves doing a full scan on the materialized table, so we call it "Materialization-scan".

If you run a join from Countries to the materialized table, the cheapest way to find a match in the materialized table is to make a lookup on its primary key (it has one: we used it to remove duplicates). Because of that, we call the strategy "Materialization-lookup".

Semi-join materialization in action

Materialization-Scan

If we chose to look for cities with a population greater than 7 million, the optimizer will use Materialization-Scan and EXPLAIN will show this:

Here, you can see:

There are still two SELECTs (look for columns with id=1 and id=2)
The second select (with id=2) has select_type=MATERIALIZED. This means it will be executed and its results will be stored in a temporary table with a unique key over all columns. The unique key is there to prevent the table from containing any duplicate records.

The optimizer chose to do a full scan over the materialized table, so this is an example of a use of the Materialization-Scan strategy.

As for execution costs, we're going to read 15 rows from table City, write 15 rows to materialized table, read them back (the optimizer assumes there won't be any duplicates), and then do 15 eq_ref accesses to table Country. In total, we'll do 45 reads and 15 writes.

By comparison, if you run the EXPLAIN with semi-join optimizations disabled, you'll get this:

...which is a plan to do (239 + 239*15) = 3824 table reads.

Materialization-Lookup

Let's modify the query slightly and look for countries which have cities with a population over one millon (instead of seven):

The EXPLAIN output is similar to the one which used Materialization-scan, except that:

the <subquery2> table is accessed with the eq_ref access method
the access uses an index named distinct_key

This means that the optimizer is planning to do index lookups into the materialized table. In other words, we're going to use the Materialization-lookup strategy.

With optimizer_switch='semijoin=off,materialization=off'), one will get this EXPLAIN:

One can see that both plans will do a full scan on the Country table. For the second step, MariaDB will fill the materialized table (238 rows read from table City and written to the temporary table) and then do a unique key lookup for each record in table Country, which works out to 238 unique key lookups. In total, the second step will cost (239+238) = 477 reads and 238 temp.table writes.

Execution of the latter (DEPENDENT SUBQUERY) plan reads 18 rows using an index on City.Country for each record it receives for table Country. This works out to a cost of (18*239) = 4302 reads. Had there been fewer subquery invocations, this plan would have been better than the one with Materialization. By the way, MariaDB has an option to use such a query plan, too (see ), but it did not choose it.

Subqueries with grouping

MariaDB is able to use Semi-join materialization strategy when the subquery has grouping (other semi-join strategies are not applicable in this case).

This allows for efficient execution of queries that search for the best/last element in a certain group.

For example, let's find cities that have the biggest population on their continent:

the cities are:

Factsheet

Semi-join materialization

Can be used for uncorrelated IN-subqueries. The subselect may use grouping and/or aggregate functions.
Is shown in EXPLAIN as type=MATERIALIZED for the subquery, and a line withtable=<subqueryN> in the parent subquery.
Is enabled when one has both materialization=on and semijoin=on

_{This page is licensed: CC BY-SA / Gnu FDL}