Trailing extended request deadlocks after pipeline splitting

[apkipa]

PgDog version

v0.1.57 (main 7594279; session and transaction modes)

Description

A legal extended-protocol pipeline containing two executed queries, a portal Describe, and one Sync completes directly but stalls through PgDog after the first two results. split_extended() inserts Flush after each Execute, then emits the trailing portal Describe and original Sync as separate requests. PostgreSQL does not send the Describe response until Flush or Sync (pq_flush()), but client_messages() waits for that response before it processes the separated Sync. The wait survives frontend close, leaving the backend checked out in both tested pool modes; repeated requests can exhaust the affected pool.

Reproduce

Install psycopg[binary] and run:

#!/usr/bin/env python3
import time

from psycopg.pq import ExecStatus, PGconn


TIMEOUT = 3.0


def run(label: str, dsn: str) -> None:
    conn = PGconn.connect(dsn.encode())
    try:
        if conn.status != 0:
            raise RuntimeError(conn.error_message.decode(errors="replace"))

        conn.enter_pipeline_mode()
        conn.send_query_params(b"SELECT 1", None)
        conn.send_query_params(b"SELECT 2", None)
        conn.send_describe_portal(b"")
        conn.pipeline_sync()

        statuses = []
        deadline = time.monotonic() + TIMEOUT
        while time.monotonic() < deadline:
            conn.consume_input()
            if not conn.is_busy():
                result = conn.get_result()
                if result is None:
                    if statuses and statuses[-1] == "PIPELINE_SYNC":
                        break
                    time.sleep(0.01)
                    continue
                statuses.append(ExecStatus(result.status).name)
            else:
                time.sleep(0.01)
            if statuses and statuses[-1] == "PIPELINE_SYNC":
                break

        print(f"{label}: statuses={statuses}")
    finally:
        try:
            conn.finish()
        except Exception:
            pass


run("direct", DIRECT_DSN)
run("pgdog", PGDOG_DSN)

Expected behavior

Both endpoints should print:

direct: statuses=['TUPLES_OK', 'TUPLES_OK', 'COMMAND_OK', 'PIPELINE_SYNC']
pgdog: statuses=['TUPLES_OK', 'TUPLES_OK', 'COMMAND_OK', 'PIPELINE_SYNC']

Actual behavior

The reproducer prints:

direct: statuses=['TUPLES_OK', 'TUPLES_OK', 'COMMAND_OK', 'PIPELINE_SYNC']
pgdog: statuses=['TUPLES_OK', 'TUPLES_OK']

The trailing Describe response and pipeline synchronization result never arrive.