Speed up database inserts from ORM - django

Question

I have a Django view which creates 500-5000 new database INSERTS in a loop. Problem is, it is really slow! I'm getting about 100 inserts per minute on Postgres 8.3. We used to use MySQL on lesser hardware (smaller EC2 instance) and never had these types of speed issues.
Details:
Postgres 8.3 on Ubuntu Server 9.04.
Server is a "large" Amazon EC2 with database on EBS (ext3) - 11GB/20GB.
Here is some of my postgresql.conf -- let me know if you need more
shared_buffers = 4000MB
effective_cache_size = 7128MB
My python:
for k in kw:
k = k.lower()
p = ProfileKeyword(profile=self)
logging.debug(k)
p.keyword, created = Keyword.objects.get_or_create(keyword=k, defaults={'keyword':k,})
if not created and ProfileKeyword.objects.filter(profile=self, keyword=p.keyword).count():
#checking created is just a small optimization to save some database hits on new keywords
pass #duplicate entry
else:
p.save()
Some output from top:
top - 16:56:22 up 21 days, 20:55, 4 users, load average: 0.99, 1.01, 0.94
Tasks: 68 total, 1 running, 67 sleeping, 0 stopped, 0 zombie
Cpu(s): 5.8%us, 0.2%sy, 0.0%ni, 90.5%id, 0.7%wa, 0.0%hi, 0.0%si, 2.8%st
Mem: 15736360k total, 12527788k used, 3208572k free, 332188k buffers
Swap: 0k total, 0k used, 0k free, 11322048k cached
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
14767 postgres 25 0 4164m 117m 114m S 22 0.8 2:52.00 postgres
1 root 20 0 4024 700 592 S 0 0.0 0:01.09 init
2 root RT 0 0 0 0 S 0 0.0 0:11.76 migration/0
3 root 34 19 0 0 0 S 0 0.0 0:00.00 ksoftirqd/0
4 root RT 0 0 0 0 S 0 0.0 0:00.00 watchdog/0
5 root 10 -5 0 0 0 S 0 0.0 0:00.08 events/0
6 root 11 -5 0 0 0 S 0 0.0 0:00.00 khelper
7 root 10 -5 0 0 0 S 0 0.0 0:00.00 kthread
9 root 10 -5 0 0 0 S 0 0.0 0:00.00 xenwatch
10 root 10 -5 0 0 0 S 0 0.0 0:00.00 xenbus
18 root RT -5 0 0 0 S 0 0.0 0:11.84 migration/1
19 root 34 19 0 0 0 S 0 0.0 0:00.01 ksoftirqd/1
Let me know if any other details would be helpful.

Answer 1

One common reason for slow bulk operations like this is each insert happening in its own transaction. If you can get all of them to happen in a single transaction, it could go much faster.

Answer 2

Firstly, ORM operations are always going to be slower than pure SQL. I once wrote an update to a large database in ORM code and set it running, but quit it after several hours when it had completed only a tiny fraction. After rewriting it in SQL the whole thing ran in less than a minute.
Secondly, bear in mind that your code here is doing up to four separate database operations for every row in your data set - the get in get_or_create, possibly also the create, the count on the filter, and finally the save. That's a lot of database access.
Bearing in mind that a maximum of 5000 objects is not huge, you should be able to read the whole dataset into memory at the start. Then you can do a single filter to get all the existing Keyword objects in one go, saving a huge number of queries in the Keyword get_or_create and also avoiding the need to instantiate duplicate ProfileKeywords in the first place.