Connection leak when using r2dbc-pool & spring-tx

[1528110566]

Bug Report

Here is my demo. Update something that not exists, throw an exception in database transaction. Two statements executed parallel using Flux.

Query query = Query.query(Criteria.where("id").is(3));
Update update = Update.update("name", "3");
List<Integer> list = new ArrayList<>();
list.add(0);
list.add(1);
for (int i = 0; i < 1000; i++) {
    Flux.fromIterable(list)
            .flatMap(item -> {
                return transactionalOperator.transactional(template.update(query, update, User.class)
                        .doOnNext(res -> {
                            if (res != 1) {
                                throw new RuntimeException("test");
                            }
                        }));
            })
            .subscribe();
}

pool-max-size is 10.
After execution, actuator shows the number of idle connections is smaller than expected, which is 1.

However, when I move .doOnNext to another block, it works as expected (attention on ) behind User.class).

Query query = Query.query(Criteria.where("id").is(3));
Update update = Update.update("name", "3");
List<Integer> list = new ArrayList<>();
list.add(0);
list.add(1);
for (int i = 0; i < 1000; i++) {
    Flux.fromIterable(list)
            .flatMap(item -> {
                return transactionalOperator.transactional(template.update(query, update, User.class))
                        .doOnNext(res -> {
                            if (res != 1) {
                                throw new RuntimeException("test");
                            }
                        });
            })
            .subscribe();
}

Versions

Windows 11
JDK 1.8.0-381
Spring-parent 2.7.13
r2dbc-mysql 0.9.3
MySQL Server 9.1.0-1.el9

Steps to reproduce

https://github.com/1528110566/r2dbc-connection-leak-demo
This is a demo that can reproduce this bug.

1. call http://localhost:8080/test
2. wait background threads are finished
3. call http://localhost:8080/actuator/prometheus
4. search r2dbc_pool_acquired_connectionsand r2dbc_pool_idle_connections. We'll find the number of connection in the pool is not as expected. This can be proved with debugging on

   r2dbc-pool/src/main/java/io/r2dbc/pool/ConnectionPool.java

   Line 106 in bf6540e

   Mono<Connection> mono = this.connectionPool.acquire()

   
   We'will find the number of connection is actually smaller than expected.

Expected behavior/code

r2dbc_pool_acquired_connections is 0 and r2dbc_pool_idle_connections is 10 after execution.

Possible Solution

No idea.
I searched everywhere on GitHub, found lots of issues related to this problem.
I think this issue is very close to answer: pgjdbc/r2dbc-postgresql#661 , but it's too difficult to follow to me :(

[agorbachenko]

Hi, looks pretty much the same as #198

[1528110566]

@agorbachenko Yes, but I think spring-tx maybe a breakthrough point.