Making Twitter Faster

From my perspective, Twitter has a really really interesting technical problem to solve. How to store and retrieve a large amount of data really really quickly.

I am making some assumptions based on how I see twitter working. I have little information about how it is architected apart from some posts that suggests that it is running ruby on rails with MySQL?

Twitter is in the rare category where there is a very large number of data being added. There should be no updates (except to user information but there should be relatively very small amount of that). There is no need for transactionality. If I guess right, it should be a large amount of inserts and selects.

While a relational database is probably the only viable choice for the time being, I think that twitter can scale and perform better if all the extra bits of a relational database system was removed.

I love challenges like this. Technical ones are easier 😉

If I didn’t have a lifetime job, I would prototype this in a bit more depth. Garry pointed me in the direction of Hadoop. Having had a quick look at it, it can take care of the infrastructure, clustering and massive horizontal scaling requirements.

Now for the data layer on top. How to store and retrieve the data. HBase is probably a good option but doing it manually should be fairly straightforward too.

From my limited understanding of twitter, there are two key pieces of functionality, the timelines and search.

The timelines can be solved by storing each tweet as a file within a directory structure. My tweets would go into


The filename would be

For the public timeline, you just have a similar folder structure, but with the timestamp, for example, the timestamp 1236158897 would go into the following structure as a symlink


For search, pick up each word in the tweet and pop the tweet as a symlink into that folder. You could have a folder per word or follow the structure above.

/t/w/i/t/t/e/r/ OR


You would then have an application running on top with a distributed cache with an API to ease access into the data easier than direct file access. Running on Linux, the kernel will take care of the large part of the automatic caching and buffering as long as there is enough RAM on the box.

This can in theory be done without Hadoop in between and separating the directory structures across multiple servers but that can have complications of its own, especially with adding and removing boxes for scalability.

You are also likely to run into issues with the number of files / sub-directories limits but they can be solved by ‘archiving’ – multiple options for that too…

Thinking about this problem brought me back to the good old days of working on the search mechanism within We needed the site to deal with a large number of searches on limited hardware when the project was still classified as a pilot.

With some hard work and experimentation, we were able to reduce the search time to a tenth of the original time.

I’ll admit that I don’t know the details or the intricacies of the requirements that twitter has. I have probably over-simplified the problem but it was still fun to think about. If you can think of problems with this – let me know; I wanna turn them into opportunities 😉

  • By Ravi jagannathan, 5 March, 2009 @ 18:21

    The big assumption you are making is that the directory search within the file system has no cost.

  • By Shri, 6 March, 2009 @ 19:52

    No, I am making the assumption that directory search is less costly than an index scan.

    There are further ways to speed this up by separating the content in each of the folders.

    In reality, it would simply be making a filesystem based index.

    Without having the time to run some tests, I would not able to suggest that it would be faster than using a database system such as HBase. However, I believe it deserves a test-run…

  • By Kevin Clark, 9 March, 2009 @ 07:59

    Hi Shri,

    Interesting read i must say. What about the caching solution, does it has to be in-proc or out-proc. I know in-proc would be much faster but than do you think with application like twitter we would need more reliability by using a out-proc caching solution?


  • By Shri, 9 March, 2009 @ 11:15

    It will definitely need an out of process clustered cache. It could probably also do with an in-process ‘first level’ cache.

    It would need some benchmarking and performance testing to identify the performance boost gained though…


Other Links to this Post

WordPress Themes