Django supports persistent connections per process as described in the Databases documentation. This may use a lot of memory depending on the application configuration, e.g. how many servers are in use and how many processes are permitted per server. However, this provides maximum performance while memory lasts.
Oracle Database Resident Connection Pooling (DRCP) allows all of these processes to share a limited number of pre-allocated connections allocated by the database server. It is important however to understand that if more connections are attempted than are supported by the pool, then some connections will have to wait. However, if a process started by a WSGI application server is idle, it may not need a connection. There is still a performance advantage in having that process already started and ready to receive requests.
For more information, take a look at these links:
The test environment producing the results below is as Follows:
- CentOS 6
- Python 3.4.2 (rh-python34 from CentOS Software Collections)
- cx_Oracle 5.2.1
- Oracle 11.2.0.4
The system is a VM on a VmWare vSphere cluster - but the important thing is how much overhead DRCP adds.
The benchmark accessed a single URL on the server that:
- Does a SELECT on the database by an indexed column
- Formats an XML file using a Django template
The results are collected using Apache Bench something like this (the URL is modified):
ab -n 1000 -c 8 'https://example.com/testing/123'
The results are summarized below:
| Test Description | Requests/sec | Mean Response Time (ms) | 50% | 66% | 75% | 80% | 85% | 90% | 98% | 99% | 100% |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Connection per Request | 35.36 | 226.215 | 135 | 175 | 305 | 371 | 512 | 560 | 708 | 740 | 1210 |
| Persistent Connection (CONN_MAX_AGE) | 156.32 | 51.178 | 49 | 55 | 60 | 62 | 69 | 71 | 80 | 96 | 269 |
| Connection Pool | 139.44 | 57.371 | 56 | 63 | 67 | 70 | 75 | 79 | 87 | 93 | 265 |
While your results may vary, the point of the numbers above is that Oracle DRCP and this package add a minimal amount of overhead while also allowing multiple applications and application servers to share a pre-allocated number of connections.