The other day I was tasked with adding tagging functionality to one of the projects I’m working on. While this is quite basic functionality it comes with a few challenges - from properly configuring relations in the DB and the ORM, through properly handling operations on tags to properly handling connections between tags and tagged entities.
How tagging (usually) works?
If I were writing a blogging platform I’d do the following to implement tagging posts:
tagstable to store all the tags,
posts_to_tagstable to store connections between posts and tags,
- Configure the ORM to see many-to-many relationship between posts and tags,
- Write application code to handle tags and post-to-tags connections.
To do this right, I’d have to put a lot of consideration and care into managing tags (e.g. handling duplicate tags) and post-to-tags connections (e.g. removing the existing connections when updating a post).
When displaying posts I’d have to issue a separate query to load tags for each post I’d like to display. That’d require more resources for my app and put the DB server(s) under heavier load.
When searching for posts by tag I’d have to issue quite an expensive query with two joins in it.
Caching posts with their tags and search by tag results would require writing additional glue code in the application.
There must be an easier way to achieve this.
How tagging could work while being simple?
How about saving tags with every post? Here’s where PostgreSQL’s
type comes into play. To add tags to posts I’d just add
tags column to
posts table and set it to store an array of strings. Then, when saving a
post I’d simply take whatever tags the user entered and save them in the
When displaying posts I’d simply display the array of tags stored in the column.
When searching for posts by tag I’d simply add a
WHERE condition to the
Caching posts and search by tag results would be no different from caching any other query results.
Sounds nice and easy but would it work? How would it affect the performance?
Relational tags vs array based tags - the test
I decided to test the performance of both solutions.
NOTE: This test has nothing to do with real-world benchmark. Please do not treat what you’re going to read as an ultimate answer to the performance difference question. I’m yet to use PostgreSQL array columns in production.
The test setup
I created a few tools to set up and populate the tables with some pseudorandom data. The idea wass to create 10000 posts and tag each one with 5 randomly chosen tags. I also created three posts which I tagged with pre-defined tags to make testing queries simpler. You can download the tools from this gist.
The test queries
Test query for relational tags:
select posts.* from tags left join posts_to_tags on posts_to_tags.tag_id = tags.id left join posts on posts.id = posts_to_tags.post_id where tags.name = 'lorem' order by posts.id asc;
Test query for array based tags:
select * from flat_posts where tags @> ARRAY['lorem'] order by id asc;
The test environment and procedure
I performed the test using PostgreSQL 9.2.4 running on my MacBook Pro (2 x i5 at 2.3GHz, 8GB RAM, Samsung 830 SSD, OS X 10.8.2). Since the MacBook is my primary work and play machine, the OS environment wasn’t exactly benchmark-friendly. In fact, it wasn’t benchmark friendly at all since while testing the performance I had a number of other apps and services running :).
I tested the performance using the following procedure:
- Stop the PostgreSQL server,
- Start the server,
- Execute the query and take a note of its execution time,
- Repeat step 6 fourteen more times,
- Calculate the mean execution time.
The test results
Median query execution time for relational tags: 17.613 ms
Median query execution time for array based tags: 7.891 ms
Turns out array based tags performed better in my totally unreliable test. To be honest, I was a bit suprised when I saw how good the array filtering performance was. I expected the result to be quite the opposite.
ARRAY data type in PostgreSQL is a nice thing to have and to use. If I were
to implement the tagging functionality I mentioned in the beggining of this
post again I’d use an array columns. If I were to write, say, an Instagram
clone I’d go with different tagging implementaion. The thing is that array
column would work until there’s a small number of entries. In the tests I
performed while playing around I found that 20+ items in the tags array
caused a major performance hit.
Discuss this post on Reddit.