At work I was looking at improving the performance of one of our slower web pages. It can be rewarding to find a little piece of code that can be easily optimized. This time there were several functions that were adding 10+ sec to the page in worst case. It wasn't a problem for most clients, but when clients with who are related to many other clients hit the page they'd experience terrible performance. Here's pseudo code for the combination of anti-patterns that caused the problem:

# Projects have users and users are in different organizations
# (project can contain multiple organization's users)
activeOrganizationProjectUsers = [x for x in project.users
        if x.active and x.organization == organization]
if activeOrganizationProjectUsers:
    # do something *NOT* using activeOrganizationProjectUsers

There are two main problems with this code:

  1. It ignores the fact the project, users, and organization are backed by an ORM
  2. The list comprehension is being used to find all matching elements when only a single element is needed.

Ignoring the ORM

The code above wouldn't be too bad if these were just lists of objects in memory. But being objects that are instantiated by an ORM a number of database queries will be issued. In this particular case (w/o eager loading across user to the organization table) the following queries were executed:

  1. Join project to user and get all users for the project's id
  2. For each user load their organization (one by one) if the user is active

So in the case where there were hundreds of users on a project there were hundreds of queries executed and hundreds of User and Organization instances were instantiated. Depending on the size of the objects (and the ORM's behavior) it can take "real time" to fetch and instantiate all these large objects.

This code base has this kind of code sprinkled through out it. At one time during it's development the developers were encouraged to treat ORM backed objects as though they were Plain Old Python Objects (POPOs). The developer wouldn't necessarily see the performance degradation using small data sets either. This is one of the reasons why I like to tail the database log (or use django-debug-toolbar if I'm using Django) to see the queries go by.

Using List Comprehensions When a Single Value is Needed

To make this situation worse, the activeOrganizationProjectUsers list wasn't actually used. This is a combination of a Python anti-pattern and the ORM anti-pattern. What was required was to determine if a single active organization user existed.

I believe the original developer(s) used the list comprehension solution in a combination of ignorance and syntactic sugar. They didn't want to write a new function to do the query and put it in the User class so they used the existing class's API. The syntactic sugar was using the list comprehension to get more values than the one that was needed. If this wasn't a (potentially) expensive ORM backed operation the original code could have been:

activeOrganizationProjectUsers = False
for x in project.users:
    if x.active and x.organization == organization:
        activeOrganizationProjectUsers = True
        break
if activeOrganizationProjectUsers:
    # do something

But this solution could still query all possible user/organization combinations. The other question would be: which set is larger the organization users or the project users? It is likely looping over the organization's users looking for active ones would be more efficient anyway.

Remember the Underlying Representation

When performance matters remembering the objects are ORM backed is important. So in this case a single query was all that was required (SqlObject pseudo syntax):

activeOrganizationProjectUsers = Users.selectBy(project=project,
                                                active=True,
                                                organization=organization).count() > 0

If abstracting out the ORM's methods is important this new function could be added to the appropriate class as a method. In my case making a change to use a query resulted in cutting the page load time by two orders of magnitude.

Comments

comments powered by Disqus