Oracle Backend with SessionPool

[halturin]

This new backend uses Oracle session pool to improve database performance in threaded environment. Overall speedup for large amounts of small queries is 2-3 times.

[halturin]

Tarball with oracle_pool db backend

[Sung-jin Hong]

Can you modify the files to conform with ​Django Patch Style?

And I think adding a new backend needs design decision. By patching the existing oracle backend and making pooling an option seems to be better. Making milestone post-1.0 since modifying or adding feature to the current backend is currently out of scope of 1.0 release.

[jbronn]

Oracle pooling patch derived from tarball.

[Ian Kelly]

Modified patch with print statements removed

[Ian Kelly]

I'm trying to run this through the test suite, and I'm seeing lots of non-deterministic test failures. When I try to run the full test suite, I get an early error where it's setting up the test database and apparently trying to reset a sequence that doesn't exist. When I try to run just the basic test, the test runs, but lots of queries return dramatically incorrect results. It's not obvious where this is coming from, but it needs to be addressed before this can be committed.

[nleschev]

We have included all changes in this patch(oracle_pool.3.diff). Also has been added 4 settings.
DATABASE_SESSION_POOL=[True/False]
DATABASE_SESSION_POOL_MIN=1
DATABASE_SESSION_POOL_MAX=3
DATABASE_SESSION_POOL_INCREMENT=1

[(none)]

Milestone post-1.0 deleted

[dbergstr@…]

Patch against revision 10578 (1.1beta). Use settings.DATABASE_OPTIONS, general code cleanup.

[Matt Boersma]

I'm still not able to get cx_Oracle's SessionPool to behave correctly for the entire Django test suite. I've tried the 4/16/09 patch here as well as my own implementations, but we quickly get test case failures that I've had a very hard time tracking down.

I have not given up--I think this should offer obvious speed improvements. I'm marking this as

[Matt Boersma]

...(continued) I'm marking this as needed for the next big release.

[dbergstr@…]

I've never looked at the test suite, nor do I have a clue how cx_oracle manages connections and transactions, but perhaps problems arise because some of the tests don't do a commit? The connection would then go back into the pool with an open transaction, and when it was next pulled out of the pool, the state in that connection would be different from what is expected. In general, using any pooled resource can expose sloppy practices in the calling code (failure to properly initialize shared memory, db connection leaks, etc.).

[Matt Boersma]

Modified patch against trunk [10832] that passes the Django test suite

[erny]

Can we have a bit of feedback about if the patch is supposed to work correctly now and it's effectivity?

[Ian Kelly]

The patch works as far as we've seen. The issues that were arising in the test suite had to do with the test setup attempting to disconnect, change the connection parameters, and reconnect, which failed because of the pool. For the same reason, I suspect it will take some work to make this compatible with the multidb branch, but otherwise it's usable.

[erny]

The patch didn't work for me for 1.0.X branch rev10865. I attach al alternative patch for this version.

[erny]

Oracle Pool for 1.0.X branch

[erny]

I found that the used SessionPools getmode parameter may no match everybody's requirements. This should be a parameter. Would be great if this really was a parameter of p.acquire()

[erny]

Bad news: I get DatabaseError: ORA-03135: connection lost contact with cx_Oracle 4.4.1 which does not drop/release/reconnect the connections. After the all connections are acquired, I get: ORA-24418: Cannot open further sessions. As it's author in the cx_Oracle mailing list suggest, a pool.drop(connection) is not enough, as it raises another exception. The complete pool may be replaced.

[edcrewe]

I have packaged up a version of this pooling backend, and made it available on pypi

see ​http://pypi.python.org/pypi/django-oraclepool

This includes some tests, and other extra little bits.
It has also been tested against 1.1 and 1.2

• Ed Crewe

[Alex Gaynor]

We are past the feature phase of 1.2.

[Jacob]

Milestone 1.3 deleted

[Aymeric Augustin]

Change UI/UX from NULL to False.

[Aymeric Augustin]

Change Easy pickings from NULL to False.

[Skylar Saveland]

I have what may be a more simple approach to the problem. I branched off of the current django master on github. I can easily remove the extraneous flake8 changes and squash into one commit for a pull request:

​https://github.com/skyl/django/compare/7732-oracle-sessionpool

It appears that the tests will pass against Oracle (takes forever). However, I'm not that familiar with the ins and outs of how database backends work. If someone could look at my patch and let me know if there is something obvious I've missed, I would appreciate it. On a simple test webapp, I'm experiencing a significant speedup with the patch.

[Anssi Kääriäinen]

It seems that allowing the use of the cx_Oracle inbuilt pooler is a good idea.

As for the implementation: is the use of class level pool variable really safe? It seems the pool instance is connection settings dependent, so having just one pool per project seems suspicious.

Also, the close() implementation should probably do a rollback before releasing the connection back to the pool.

[Skylar Saveland]

I changed to 1 pool per alias; 1 pool per project was definitely wrong. I also changed to doing a self.connection.rollback() in close before releasing the connection.

[erny]

I had the following problems in the past: The Oracle pools is a "per process" pool. Each Python Process gets its own pool. Supposing X processes (apache prefork, wsgi, etc.) with each Y min connections we get X * Y connections. So keep your min connections count low. With Oracle 11, there seems to be a server side session pool which should be much more convenient, but I didn't test it.

[Skylar Saveland]

Unfortunately, I think there needs to be some error-checking on the connection after it is acquired from the pool. After some hours (I didn't configure the instance I'm running against so I'm not sure how long), I get ORA-02399: exceeded maximum connect time, you are being logged off; then, the pool continues to try to get the same connection, failing every time with, ORA-01012: not logged on. I think there are 2 options then.

1) put a CONN_MAX_AGE parameter in the SESSION_POOL settings dict and drop the connection if it is close enough to that age
2) ping the connection after it is acquired and drop it if it is not usable

(1) Age is not an attribute of the connection that I can see; so, it would be ugly to do our own book-keeping on the age of the connections in the pools we are holding. This does not cover other possible error scenarios.
(2) This would be a hit to performance every time a connection is acquired.

I think (2) would be easy enough to add. Other suggestions?

[Omer Katz]

Since I believe that adding pooled backends was rejected from Django shouldn't this issue be closed?

[Dan Davis]

I'm interested in this. I'm not quite on the point of assigning it, but one of my developers asked me see to return to what is needed to use Oracle DRCP. I tested it using django-oraclepool many years ago to determine whether there was any performance difference, from the client-side, between CONN_MAX_AGE and this feature. django-oraclepool is so old, that it seems I'd either need to fork it or return to this issue.