19.1. Logical Planner
Presto currently has a rule based optimizer. A given SQL query goes through syntactic analysis, semantic analysis, logical planning, physical planning and then execution. Syntactic analysis, or parsing, is used to break down the query into its component structures and check that the query conforms to the syntactic rules of SQL. The next step is semantic analysis, which gathers the semantic information from the query that is necessary for planning. It deals with things like type checking and verifying that the column names used in the query exist for a given table. The logical planner then takes this syntactic and semantic information for a query and using metadata and session information creates a base logical plan.
Consider this example:
SELECT nationkey FROM nation JOIN region ON nation.regionkey = region.regionkey WHERE nationkey > 9
For this query Presto will generate a logical base plan:
ProjectNode | FilterNode | JoinNode / \ ProjectNode ProjectNode | | TableScanNode TableScanNode (nation) (region)
The base logical plan is then optimized using a list of optimizer transformation rules like predicate push down or pruning of unused columns to get the final logical plan. For example the optimization predicate push down ensures that the predicates are applied as early in the plan as possible to reduce the amount of data that has to be processed. In the given query, the filter on nationkey column will be pushed below the JoinNode and the ProjectNode towards the TableScanNode for nation table and the transformed plan will look like this:
ProjectNode | JoinNode / \ ProjectNode ProjectNode | | FilterNode TableScanNode | (region) TableScanNode (nation)
The final optimized logical plan is then fragmented into sub-plans during physical planning. These fragments are shipped for execution to distributed nodes.