rest: too many open connections to database

[mdonadoni]

Many connections to the DB are opened by r-server when handling requests. These connections are not closed, probably because of connection pooling.

Given that there is a limit on the number of concurrent connections, it can happen that under heavy load some components are not able to reach the DB:

==> ERROR: Cannot create workflow :
(psycopg2.OperationalError) connection to server at "..." (...), port ... failed: FATAL:  remaining connection slots are reserved for non-replication superuser connections

How to reproduce:

1. Deploy REANA locally in production mode

2. Execute many workflows and send many requests to REANA, like so:

   $ for i in `seq 30`; do reana-client list &; done; wait
   $ for i in `seq 30`; do reana-client ls -w workflow &; done; wait

3. Look at open connections (in this case 10.244.0.178 is r-server, while 10.244.0.5 is r-w-controller)

   reana=# select client_addr, count(*) from pg_stat_activity group by client_addr;
    client_addr  | count 
   --------------+-------
                 |     6
    10.244.0.178 |    46
    10.244.0.5   |     5
   (3 rows)
   

[mdonadoni]

This issue is (at least partially) caused by Invenio and REANA using different connection pools to connect to the DB.

This can be seen easily by setting the following reana-server's environment variables in values.yaml:

@@ -78,6 +78,8 @@ components:
       REANA_RATELIMIT_GUEST_USER: "20 per second"
       REANA_RATELIMIT_AUTHENTICATED_USER: "20 per second"
       REANA_RATELIMIT_SLOW: "1/5 second"
+      SQLALCHEMY_POOL_SIZE: 1
+      SQLALCHEMY_MAX_OVERFLOW: 0
     uwsgi:
       processes: 6
       threads: 4

In this case, the pool size used by REANA and Invenio is different, which means that different connection pools are being used:

$ kubectl -it exec deployment/reana-server -c rest-api -- invenio shell
In [1]: from invenio_db import db

In [2]: db.engine.pool.size()
Out[2]: 5

In [3]: from reana_db import database

In [4]: database.engine.pool.size()
Out[4]: 1

Note that Invenio uses Flask-SQLAlchemy extension to connect to the DB, so it is possible to modify Invenio's pool size by setting SQLALCHEMY_POOL_SIZE = 1 in reana_server/config.py.

At the moment I don't see which are the benefits of using different connection pools. Instead, there a few drawbacks to this approach: not only it increases the number of connections open with the DB, but it is also possible that multiple connections to the DB are opened to serve a single request. This also makes it impossible to modify REANA and Invenio tables in the same transaction.

[mdonadoni]

This is the number of connections depending on the number of uWSGI processes and threads.

With REANA's and Invenio's pool size equal to 1:

Processes	Threads	Connections
3	4	7 (3 REANA + 3 Invenio + 1 scheduler)
3	8	7 (3 + 3 + 1)
6	4	13 (6 + 6 + 1)
6	8	13 (6 + 6 + 1)

With REANA's pool size set to 1 and Invenio's pool size equal to 5:

Processes	Threads	Connections
3	4	16 (3 REANA + 3*4 Invenio + 1 scheduler)
3	8	19 (3 + 3*5 + 1)
6	4	31 (6 + 6*4 + 1)
6	8	37 (6 + 6*5 + 1)

Connections are shared between threads but not between processes. For this reason, it is not very useful to have a pool size bigger than the number of threads per process, as each thread will open one connection and none of them will be shared (e.g. with pool size equal to 5 and considering 4 threads, only 4 connections will be opened).

[mdonadoni]

Note that REANA now supports pgBouncer, so this issue has become less urgent