Penyaringan Partisi saat Perencanaan dan Eksekusi

You partitioned a huge table so PostgreSQL can skip partitions that cannot contain matching rows. That skipping is called partition pruning.

Without pruning, a query that touches one month of data could still scan every partition for every month. The whole performance benefit of partitioning depends on the planner (and sometimes the executor) recognising which partitions are relevant.

• Plan-time pruning happens when the planner already knows the filter values.
• Execution-time pruning happens when the values are only known once the query runs.

In this lesson you will learn to confirm both kinds by reading EXPLAIN output.

Throughout the lesson we use a range-partitioned events table, partitioned by month on created_at. Each child partition holds one month of rows.

This is the classic time-series layout where pruning pays off the most: queries usually target a narrow date range, so most partitions should be skipped entirely.

When the filter compares the partition key to a constant, the planner can decide which partitions to touch before execution. This is static (plan-time) pruning.

Run EXPLAIN on a query restricted to one month. The plan should reference only the matching partition(s); the others never appear.

The setting enable_partition_pruning (on by default) controls this behaviour.

Here is the shape of the plan for that single-month query. Notice only events_2024_02 is scanned. events_2024_01 and events_2024_03 are absent from the plan entirely.

• The Append (or Seq Scan with one child) lists only surviving partitions.
• Pruned partitions leave no trace in the plan output.

This is the clearest confirmation of static pruning: count the partitions in the plan and compare against how many exist.

If your EXPLAIN shows an Append over all partitions, static pruning did not apply. Common causes:

• The filter does not reference the partition key (e.g. filtering on user_id only).
• A function wraps the key, e.g. date(created_at) = '2024-02-10', hiding the relationship from the planner.
• The value is not a constant at plan time (parameter or join column) — that needs execution-time pruning.

Keep the partition key bare on one side of the comparison so the planner can match it to partition bounds.

With a prepared statement or a value supplied at run time, the planner may not know the constant when it builds the plan. A generic plan must stay valid for any parameter value, so it cannot statically prune.

Instead PostgreSQL defers the decision: it keeps all partitions in the plan but adds the ability to skip them while the query executes. This is execution-time (runtime) pruning.

Runtime pruning advertises itself in EXPLAIN with two key markers under an Append node:

• Subplans Removed: N — partitions discarded before scanning.
• For parameterised plans, a line like Filter or initplan parameters that drive the pruning.

If you see Subplans Removed, the executor pruned partitions at run time. If you see neither that nor a reduced partition list, no pruning occurred.

Execution-time pruning is not only for parameters. It also kicks in when the partition key is compared to a value produced by the outer side of a join (a Nested Loop), or by a subquery.

Each outer row supplies a key value, and for each one the executor prunes down to the relevant partition. This is extremely valuable for selective lookups against a partitioned fact table.

Plain EXPLAIN shows what could be pruned. To prove pruning happened during a real run — especially runtime pruning — use EXPLAIN ANALYZE.

• Subplans Removed: N appears with the actual number removed at execution.
• Surviving partitions show actual rows; pruned ones show (never executed) if they remain as subnodes.

Reading actual time and actual rows per partition tells you exactly which children did work.

Older PostgreSQL relied on constraint_exclusion for inheritance-based partitioning. Modern declarative partitioning uses partition pruning, which is faster and supports runtime pruning.

• enable_partition_pruning = on drives the new mechanism (plan and execution time).
• constraint_exclusion only ever worked at plan time and only with CHECK constraints.

For declarative partitions, leave enable_partition_pruning on and do not depend on constraint_exclusion.

When verifying pruning on a real query, work through this list:

• Is the partition key in the predicate, bare and SARGable? No wrapping functions, no implicit casts that block matching.
• Static case: run EXPLAIN — do pruned partitions disappear from the plan?
• Runtime case: look for Subplans Removed: N under Append.
• Confirm with EXPLAIN ANALYZE that only the expected partitions did work.
• Check the setting if nothing prunes: enable_partition_pruning must be on.

You can now confirm partition pruning from EXPLAIN output:

• Static pruning happens at plan time when the partition key is compared to constants; pruned partitions simply vanish from the plan.
• Execution-time pruning handles parameters and join-driven values; look for Subplans Removed: N under Append.
• Keep the partition key bare and SARGable — wrapping it in a function blocks pruning.
• Use EXPLAIN ANALYZE to prove which partitions actually did work, and verify enable_partition_pruning is on when nothing prunes.

Reading these markers turns partitioning from a hopeful design into a verified performance win.