带有 Django 的 CONN_MAX_AGE 的 pgbouncer 的理想设置

Question

我正在运行一个多租户网站，我想减少为每个请求创建 PostgreSQL 连接的开销。 Django 的 CONN_MAX_AGE 允许这样做，但代价是创建大量到 PostgreSQL 的开放空闲连接（8 个工作人员 * 20 个线程 = 160 个连接）。每个连接 10MB，这会消耗大量内存。

主要目的是减少连接时间开销。 因此我的问题是：

• 我应该为这样的解决方案使用哪种设置？ （PgBouncer？）
• 我可以在 Django 中使用“事务”池模式吗？
• 使用https://github.com/kennethreitz/django-postgrespool 代替 Django 的池会更好吗？

Django 1.6 设置：

DATABASES['default'] = {
    'ENGINE':   'django.db.backends.postgresql_psycopg2',

     ....

    'PORT': '6432'
    'OPTIONS': {'autocommit': True,},
    'CONN_MAX_AGE': 300,
}

ATOMIC_REQUESTS = False   # default

Postgres：

max_connections = 100

PgBouncer：

pool_mode = session     # Can this be transaction?
max_client_conn = 400   # Should this match postgres max_connections?
default_pool_size = 20
reserve_pool_size = 5

Answer 1

这是我使用的设置。

pgbouncer 与 gunicorn、celery 等在同一台机器上运行。

pgbouncer.ini：

[databases]
<dbname> = host=<dbhost> port=<dbport> dbname=<dbname>

[pgbouncer]
: your app will need filesystem permissions to this unix socket
unix_socket_dir = /var/run/postgresql
; you'll need to configure this file with username/password pairs you plan on
; connecting with.
auth_file = /etc/pgbouncer/userlist.txt

; "session" resulted in atrocious performance for us. I think
; "statement" prevents transactions from working.
pool_mode = transaction

; you'll probably want to change default_pool_size. take the max number of
; connections for your postgresql server, and divide that by the number of
; pgbouncer instances that will be conecting to it, then subtract a few
; connections so you can still connect to PG as an admin if something goes wrong.
; you may then need to adjust min_pool_size and reserve_pool_size accordingly.
default_pool_size = 50
min_pool_size = 10
reserve_pool_size = 10
reserve_pool_timeout = 2
; I was using gunicorn + eventlet, which is why this is so high. It
; needs to be high enough to accommodate all the persistent connections we're
; going to allow from Django & other apps.
max_client_conn = 1000
...

/etc/pgbouncer/userlist.txt：

"<dbuser>" "<dbpassword>"

Django 设置.py：

...
DATABASES = {
    'default': {
        'ENGINE': 'django.contrib.gis.db.backends.postgresql_psycopg2',
        'NAME': '<dbname>',
        'USER': '<dbuser>',
        'PASSWORD': '<dbpassword>',
        'HOST': '/var/run/postgresql',
        'PORT': '',
        'CONN_MAX_AGE': None,  # Set to None for persistent connections
    }
}
...

如果我没记错的话，你基本上可以有任意数量的“持久”连接到 pgbouncer，因为当 Django 完成它们时，pgbouncer 会将服务器连接释放回池中（只要你使用 transaction 或 @ 987654326@ 为pool_mode）。当 Django 尝试重用其持久连接时，pgbouncer 负责等待与 Postgres 的可用连接。