From: Ilmar Yunusov <tanswis42@gmail.com>
To: pgsql-hackers@postgresql.org
Cc: Ilmar Yunusov <tanswis42@gmail.com>
Subject: [RFC PATCH v0 0/7] Add EXPLAIN ANALYZE wait event reporting
Date: Sat,  9 May 2026 04:22:30 +0500
Message-ID: <cover.1778280923.git.tanswis42@gmail.com>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Archived-At: 
 <https://www.postgresql.org/message-id/cover.1778280923.git.tanswis42%40gmail.com>
Precedence: bulk

This RFC prototype adds `EXPLAIN (ANALYZE, WAITS)`, which reports
completed wait intervals observed through `pgstat_report_wait_start/end()`.
The option is named `WAITS` in this RFC to match the short style of
`BUFFERS`, `WAL`, `IO`, and `MEMORY`.  I am not attached to the exact name;
`WAIT_EVENTS` may be clearer but is more verbose.

PostgreSQL already exposes a backend's current wait event through
pg_stat_activity.  This patch explores making the same wait event
instrumentation useful in EXPLAIN ANALYZE by collecting per-statement and
per-plan-node wait event usage while a statement executes.

Statement-level output is reported as `Statement Wait Events`.  It counts
each completed wait once per active statement-level collector and includes
parallel worker waits.  Nested EXPLAIN ANALYZE WAITS collectors maintain
separate statement-level summaries; a wait is counted once in each active
collector.

Plan-node output is reported as `Wait Events`.  Node-level attribution is
intentionally inclusive, matching EXPLAIN ANALYZE node timing: a wait is
attributed to every active plan node captured when the wait begins.  This
means parent and child nodes can show the same wait, and node-level wait
times must not be summed to compute a statement total.

The implementation keeps wait-end accounting allocation-free.  Each
statement and plan-node accumulator preallocates storage for 64 distinct
wait event identities; additional distinct identities are accumulated in
`Unrecorded Wait Event Calls` and `Unrecorded Wait Event Time` without event
identity.  The fixed bound is intended to make the wait-end path predictable
and safe in places where allocation would be undesirable.  The overflow
bucket preserves total calls/time, but loses per-event identity.  This is a
deliberate RFC point.

Patch layout:

1. add statement-level EXPLAIN WAITS reporting;
2. aggregate statement-level waits from parallel workers;
3. add plan-node wait attribution, including manual executor paths;
4. refine attribution semantics, docs, overflow output, and tests;
5. harden accumulator handling and keep wait-end allocation-free;
6. hide accumulator internals behind the wait-event accounting API;
7. update EXPLAIN option tab completion.

Important review questions:

- Is the `WAITS` option name and output shape acceptable, or should this be
  `WAIT_EVENTS` / different labels?
- Is inclusive per-node attribution the right semantic for EXPLAIN?
- Is the fixed 64-entry accumulator plus explicit overflow bucket acceptable?
- Is the disabled hot-path overhead of checking an exported boolean in
  pgstat_report_wait_start/end acceptable?
- Are the test scaffolding choices acceptable, especially planner GUCs and
  pg_sleep wrappers used to force deterministic wait-attribution cases?  The
  tests use pg_sleep only to force a stable Timeout:PgSleep wait identity;
  durations are normalized by the existing EXPLAIN test filters.

Local verification so far:

- `make -s -j4`
- `make -C doc/src/sgml check`
- `make -s -C src/bin/psql`
- `make -C src/test/regress check-tests TESTS='test_setup create_index explain'`
- `git diff --check`

The final diff of this 7-patch branch is identical to the development branch
`r314tive/pg-wait-explain-mvp`.

Local optimized macOS microbenchmarks are directional only.  The current
synthetic C wait-loop run measured roughly 0.1-0.2 ns/wait disabled overhead
and about 30 ns/wait enabled accounting for a single active node.  These
numbers are not intended as performance evidence for commit; they only served
as a local smoke check that the disabled path is plausibly small.  I would
want repeated Linux, CPU-pinned numbers before drawing stronger conclusions.

Ilmar Yunusov (7):
  Add EXPLAIN WAITS statement reporting
  Aggregate EXPLAIN WAITS from parallel workers
  Attribute EXPLAIN WAITS to plan nodes
  Refine EXPLAIN WAITS attribution semantics
  Harden EXPLAIN WAITS accumulator handling
  Hide EXPLAIN WAITS accumulator internals
  Keep EXPLAIN option completion current

 doc/src/sgml/ref/explain.sgml              |  61 ++++
 src/backend/commands/explain.c             | 172 +++++++++-
 src/backend/commands/explain_state.c       |   8 +
 src/backend/executor/execAsync.c           |  22 ++
 src/backend/executor/execMain.c            |   1 +
 src/backend/executor/execParallel.c        | 295 ++++++++++++++++-
 src/backend/executor/execProcnode.c        |  24 +-
 src/backend/executor/execUtils.c           |   1 +
 src/backend/executor/instrument.c          |   7 +
 src/backend/executor/nodeBitmapAnd.c       |   7 +
 src/backend/executor/nodeBitmapIndexscan.c |   7 +
 src/backend/executor/nodeBitmapOr.c        |   7 +
 src/backend/executor/nodeHash.c            |   7 +
 src/backend/utils/activity/wait_event.c    | 363 +++++++++++++++++++++
 src/bin/psql/tab-complete.in.c             |   6 +-
 src/include/commands/explain_state.h       |   1 +
 src/include/executor/execParallel.h        |   2 +
 src/include/executor/instrument.h          |   1 +
 src/include/nodes/execnodes.h              |   3 +
 src/include/utils/wait_event.h             |  45 +++
 src/test/regress/expected/explain.out      | 202 ++++++++++++
 src/test/regress/sql/explain.sql           | 144 ++++++++
 22 files changed, 1371 insertions(+), 15 deletions(-)

-- 
2.52.0