Description

PgSingleTServerTest.ManyYsqlConnections fails on both Linux and macOS due to a self-deadlock in
the kNormal PgClient RPC pool.

PgClientService::TriggerRelcacheInitConnection is a synchronous RPC handler. It schedules
MakeRelcacheInitConnection on the messenger's scheduler and then calls future::wait_for on the
in-flight promise, parking the kNormal worker. MakeRelcacheInitConnection itself runs on
kNormal, and opens a libpq connection to the local PG listener whose backend issues PgClient RPCs
that also land on kNormal.

Under enough concurrent relcache-init callers (e.g. ManyYsqlConnections with 64 backends and a
32-thread kNormal pool), every kNormal worker parks in wait_for, and the only thing that
could call promise.set_value can never get a worker — classic self-deadlock.

Steps to reproduce

yb_build.sh --cxx-test pgwrapper_pg_single_tserver-test \
  --gtest_filter PgSingleTServerTest.ManyYsqlConnections

Affects

• Linux (release + debug)
• macOS

Issue Type

kind/failing-test

Warning

I confirm this issue does not contain any sensitive information.