Because each worker executes the parallel portion of the plan to completion, it is not possible to simply take an ordinary query plan and run it using multiple workers. Each worker would produce a full copy of the output result set, so the query would not run any faster than normal but would produce incorrect results. Instead, the parallel portion of the plan must be what is known internally to the query optimizer as a partial plan; that is, it must be constructed so that each process which executes the plan will generate only a subset of the output rows in such a way that each required output row is guaranteed to be generated by exactly one of the cooperating processes. Generally, this means that the scan on the driving table of the query must be a parallel-aware scan.
The following types of parallel-aware table scans are currently supported.
In a parallel sequential scan, the table's blocks will be divided among the cooperating processes. Blocks are handed out one at a time, so that access to the table remains sequential.
In a parallel bitmap heap scan, one process is chosen as the leader. That process performs a scan of one or more indexes and builds a bitmap indicating which table blocks need to be visited. These blocks are then divided among the cooperating processes as in a parallel sequential scan. In other words, the heap scan is performed in parallel, but the underlying index scan is not.
In a parallel index scan or parallel index-only scan, the cooperating processes take turns reading data from the index. Currently, parallel index scans are supported only for btree indexes. Each process will claim a single index block and will scan and return all tuples referenced by that block; other process can at the same time be returning tuples from a different index block. The results of a parallel btree scan are returned in sorted order within each worker process.
Other scan types, such as scans of non-btree indexes, may support parallel scans in the future.
Just as in a non-parallel plan, the driving table may be joined to one or more other tables using a nested loop, hash join, or merge join. The inner side of the join may be any kind of non-parallel plan that is otherwise supported by the planner provided that it is safe to run within a parallel worker. Depending on the join type, the inner side may also be a parallel plan.
In a nested loop join, the inner side is always non-parallel. Although it is executed in full, this is efficient if the inner side is an index scan, because the outer tuples and thus the loops that look up values in the index are divided over the cooperating processes.
In a merge join, the inner side is always a non-parallel plan and therefore executed in full. This may be inefficient, especially if a sort must be performed, because the work and resulting data are duplicated in every cooperating process.
In a hash join (without the "parallel" prefix), the inner side is executed in full by every cooperating process to build identical copies of the hash table. This may be inefficient if the hash table is large or the plan is expensive. In a parallel hash join, the inner side is a parallel hash that divides the work of building a shared hash table over the cooperating processes.
PostgreSQL supports parallel aggregation by aggregating in
two stages. First, each process participating in the parallel portion of
the query performs an aggregation step, producing a partial result for
each group of which that process is aware. This is reflected in the plan
as a Partial Aggregate node. Second, the partial results are
transferred to the leader via Gather or Gather
Merge. Finally, the leader re-aggregates the results across all
workers in order to produce the final result. This is reflected in the
plan as a Finalize Aggregate node.
Because the Finalize Aggregate node runs on the leader
process, queries which produce a relatively large number of groups in
comparison to the number of input rows will appear less favorable to the
query planner. For example, in the worst-case scenario the number of
groups seen by the Finalize Aggregate node could be as many as
the number of input rows which were seen by all worker processes in the
Partial Aggregate stage. For such cases, there is clearly
going to be no performance benefit to using parallel aggregation. The
query planner takes this into account during the planning process and is
unlikely to choose parallel aggregate in this scenario.
Parallel aggregation is not supported in all situations. Each aggregate
must be safe for parallelism and must
have a combine function. If the aggregate has a transition state of type
internal, it must have serialization and deserialization
functions. See CREATE AGGREGATE for more details.
Parallel aggregation is not supported if any aggregate function call
contains DISTINCT or ORDER BY clause and is also
not supported for ordered set aggregates or when the query involves
GROUPING SETS. It can only be used when all joins involved in
the query are also part of the parallel portion of the plan.
Whenever PostgreSQL needs to combine rows
from multiple sources into a single result set, it uses an
Append or MergeAppend plan node.
This commonly happens when implementing UNION ALL or
when scanning a partitioned table. Such nodes can be used in parallel
plans just as they can in any other plan. However, in a parallel plan,
the planner may instead use a Parallel Append node.
When an Append node is used in a parallel plan, each
process will execute the child plans in the order in which they appear,
so that all participating processes cooperate to execute the first child
plan until it is complete and then move to the second plan at around the
same time. When a Parallel Append is used instead, the
executor will instead spread out the participating processes as evenly as
possible across its child plans, so that multiple child plans are executed
simultaneously. This avoids contention, and also avoids paying the startup
cost of a child plan in those processes that never execute it.
Also, unlike a regular Append node, which can only have
partial children when used within a parallel plan, a Parallel
Append node can have both partial and non-partial child plans.
Non-partial children will be scanned by only a single process, since
scanning them more than once would produce duplicate results. Plans that
involve appending multiple results sets can therefore achieve
coarse-grained parallelism even when efficient partial plans are not
available. For example, consider a query against a partitioned table
which can be only be implemented efficiently by using an index that does
not support parallel scans. The planner might choose a Parallel
Append of regular Index Scan plans; each
individual index scan would have to be executed to completion by a single
process, but different scans could be performed at the same time by
different processes.
enable_parallel_append can be used to disable this feature.
If a query that is expected to do so does not produce a parallel plan, you can try reducing parallel_setup_cost or parallel_tuple_cost. Of course, this plan may turn out to be slower than the serial plan which the planner preferred, but this will not always be the case. If you don't get a parallel plan even with very small values of these settings (e.g. after setting them both to zero), there may be some reason why the query planner is unable to generate a parallel plan for your query. See Section 15.2 and Section 15.4 for information on why this may be the case.
When executing a parallel plan, you can use EXPLAIN (ANALYZE,
VERBOSE) to display per-worker statistics for each plan node.
This may be useful in determining whether the work is being evenly
distributed between all plan nodes and more generally in understanding the
performance characteristics of the plan.