Elixir operations with `mode: :savepoint` leave a `wait_event='ClientRead'` transaction hanging

[kzemek]

PgDog version
v0.1.49, v0.1.51 & main

Description
When running Elixir's Postgrex and using its mode: :savepoint feature, Elixir side times out while PgDog side leaves the connection in a wait_event='ClientRead' transaction.

---

Here's the simplest repro, you can run it with elixir repro.exs:

Mix.install([{:postgrex, "~> 0.19"}])

{:ok, conn} =
  Postgrex.start_link(
    hostname: "127.0.0.1",
    port: 6432,
    username: "postgres",
    password: "postgres",
    database: "postgres"
  )

result =
	Postgrex.transaction(
	  conn,
	  fn c -> Postgrex.query(c, "SELECT 1/0", [], mode: :savepoint) end,
	  timeout: 5_000
	)

dbg(result)

then verify the hanging transaction:

SELECT pid, state, wait_event, query FROM pg_stat_activity WHERE wait_event = 'ClientRead'

---

Postgrex is doing

Query("BEGIN")
Query("SAVEPOINT postgrex_query")
Bind
Execute
Query("RELEASE SAVEPOINT postgrex_query")

it doesn't do Sync, which seems to be what trips PgDog here, as ErrorResponse handling re-queues ReadyForQuery regardless if Sync was sent. Best I can tell, what Postgrex does is unique but legal.

Logs

pgdog-trace-full.log

Configuration

The example Docker compose configuration from the repo. Tested in both sharded and unsharded modes just in case.

[levkk]

Perfect timing, we just added an Elixir test suite: #1310

I'll try to take a look at this asap.

[kzemek]

Related but tangential: despite best efforts, state bugs can happen 😄 In this case the connection never returned to a pool, so it didn't go through Guard::cleanup() checks. It would help if self.client_messages(&mut query_engine).await? could be broken by shutdown.notified() or by context.stream.check()