The Highly Scalable Architecture behind the Wajam Social Search Engine

August 6th, 2013 | by alainwong

posted in Big Data, Engineering

At Wajam, we’re helping you find recommendations from friends you trust, whenever you need them. In order to do that, we constantly innovate with new technologies. Our technical blog posts are meant to share with you what we’ve learned along our high-tech adventures.

Graffiti artwork from the Wajam office


Wajam is a social search engine that gives you access to the knowledge of your friends. We gather your friends’ recommendations from Facebook, Twitter and other social platforms and serve these back to you on supported sites like Google, eBay, TripAdvisor and Wikipedia.

To do this, we aggregate, analyze and index relevant pieces of shared information on our users’ social networks. The main challenge stems from the amount of information we collect as well as the speed at which we must display search results back to users. Our goal is to serve requests in <200ms to remain competitive.


At the beginning, Wajam searches were made in MySQL and soon after using a Sphinx cluster. Sphinx was chosen at the time because some of the first developers at Wajam had experience with it.

However, Sphinx was not suited to be distributed and real-time, so the team had to build a large framework to fulfill these requirements. With time, the solution began to show its age; we encountered low-level and painful cluster management, manual resharding, slow geolocated searches, and high CPU usage.

At the breaking point, we faced a crossroad. Either maintain the current solution by building a bigger framework around it, or go with a new modern solution.


After examining our options which included Solr, we decided to give Elasticsearch a try.

Elasticsearch is built around Apache Lucene, which is a robust and widely used search engine and has a large developer community associated to it. Elasticsearch is built with scalability and real-time indexing as a prime concern, two of the main reasons why we wanted to move over from Sphinx. In addition, Elasticsearch uses a technique called geohash for geo-located queries which offer a significant performance boost for those kinds of queries.

Another important point is the fact that Elasticsearch is written in Java while Sphinx is written in procedural C++. The advantage of well-maintained and open-source Java code is more than just stability, it also means that we could code features ourselves and give back to the project.

Here is a comparison table between Elasticsearch and Sphinx highlighting the key points according to our needs.


At the time of our evaluation in fall 2012, Elasticsearch seemed to gain traction quite rapidly and came out as an interesting alternative that could potentially fill all of our needs.

We were impressed by preliminary benchmarks using our data. At smaller scale, Elasticsearch outperformed Sphinx in most of our use case scenarios. We then decided to fully commit to the transition.


  • Scalability : Horizontal scalability is one of our prime concerns as the quantity of data generated by Wajam is growing exponentially, as is our number of users.
  • Performance : We want to improve performance on every query done to the Search API. The initial goal was to bring search latency under 200 ms (median), including query construction, query execution and parsing.
  • Flexibility : Data is often adapted to new features that we build, so we’re interested in having a flexible data scheme to suit our rapid development cycle.


Our initial idea was to develop the first iteration of the new infrastructure as a direct replacement for Sphinx. That is, keep things simple and use a single index to put in each and every document. As explained later on under pitfalls and lessons learned, this turned out to be a flawed plan.

We already had a defined API, so we built an Elasticsearch search client replacement for the Sphinx one. Following that, we defined the mapping based on what we had in Sphinx and made adjustments to make use of the new Elasticsearch features.

Realtime data comes from Jobs that fetch the data from the diverse social network APIs. The data is backed up on HDFS from the same source. Then we used a Pig UDF (Wonderdog by Infochimps) to bulk load the data into Elasticsearch.

The Backend API and the Jobs were converted to interact with Elasticsearch.

During the transition period, we were pushing data in parallel to Elasticsearch and Sphinx at the same time. To complete the migration, we progressively changed a configuration flag in the Backend to query Elasticsearch instead of Sphinx.


Elasticsearch is not Sphinx. Trying to replicate an architecture from a system to another for simplicity can be a bad idea. It might work in some cases, but in ours, it almost failed. Putting tens of billions of documents in a single index is against the Elasticsearch philosophy and recommendations. We did manage to make this work, but without replicas and some flaky performance. In this situation, each query would hit hundreds of nodes in the cluster. Theoretically this is using the cluster as its full capacity, but in practice it’s not a good idea because nodes will fail, slow down or even disappear. Losing one node means losing indexing capability and degraded search. Don’t do it.

Feature flag. Since we were doing a transition from a legacy system to a new one, we had the opportunity to build the new system in parallel and progressively switch with configuration. Since we had no previous experience with Elasticsearch and had no idea how it would react to our traffic, simply switching to the new architecture would have been risky. So we built the system with a configuration flag to switch from one search engine to another. This way we could iterate until we were able to support all the traffic or rollback at our convenience. This saved us more than once.

Monitor and track everything. Especially with systems like information retrieval, it’s hard to apprehend the impact of every change you make. Having metrics for almost everything helped us measure the impact of the iterative changes we made to our infrastructure. For this we use metrics, graphite and a nice perl script to visualize key performance indicators.

Design query (filters) that are cacheable. The cache in Elasticsearch is awesome. Elasticsearch will build the filter bitsets in memory, and continually add new segments to each bitset as they become available. From then on, query filtering is fullfilled by the cached bitsets instead of reading the segments every time, delivering better performance as a result.


Since then, we iterated many times on the architecture and we are now running a fully replicated Elasticsearch serving all of our users’ searches. We’ve gone through multiple implementations, and are currently using fixed-size monthly bucket indices to give us predictability over cluster behavior and scalability since we’re keeping equal-sized index shards. This way, we can easily model our needs according to our growth rate and add nodes when needed. Moreover, having multiple indices grouped and identified with aliases allow us to be more flexible with the queries we are doing.

During the whole process we have been monitoring the evolution with a large number of performance metrics. Below are some of the latest graphs showing the system response time and indexation rate that we have been able to achieve with the current architecture.

Median search query response time over time

Document indexing over time

Since switching over to Elasticsearch, we’ve enjoyed a 10x improvement in performance compared to Sphinx with a median response time of 150 ms while serving more documents than ever.


At Wajam, we’re always been focused on giving users the best social search experience, and now with Elasticsearch, we’re able to deliver on this promise more quickly and more reliably. You can get recommendations from friends you trust, across all of the sites that we support, both on the desktop and on mobile devices.

Although it was a challenging transition to migrate our data to the new Elasticsearch architecture, we now have a solid foundation from which to scale. Stay tuned for more exciting new features coming soon!

Elasticsearch is launching 0.90.3 today. Check it out here.

I would like to thank personally Shay Banon, the elasticsearch consulting staff and everyone on #elasticsearch on Freenode that helped me on this project. I am often online on this channel as jgagnon. You can also reach me at @jeromegagnon1.

Apache Lucene and Solr are trademarks of the Apache Software Foundation

3 Responses to “The Highly Scalable Architecture behind the Wajam Social Search Engine”

  1. Andrew says:


    > we’ve enjoyed a 10x improvement in performance compared to Sphinx with a median response time of 150 ms while serving more documents than ever.

    Uh, is that really with comparable hardware, datasets, queries, etc?

    If so, I’d be very interested at looking at a specific query type that we execute so badly.

    Because, frankly, many “X was 10x faster than Y” statements actually really mean something like “we benchmarked X with stopwords on and caching on and whatnot on too, but we never bothered to enable the indexes in Y” upon closer examination :)

    My personal favorite here is “we benchmarked a full scan w/o any indexes at all in a generic RDBMS against a custom-made compressed and indexed column store and found that the store was 100x faster” :)

    • jeromegagnon says:

      Hi Andrew,

      Thanks for the comment. Indeed, the improvements we saw were not only due to the switch in search engines, as there are many variables affecting performance. However, our main goal of switching to Elasticsearch was to ease the scalability and make use of the features that they offer like geohash for geolocated queries and automatic shard replication/distribution.

      To clarify, we added more hardware as our document base grew bigger. We also took advantage of the multi-index capacity of ElasticSearch to optimize the way we execute queries.

  2. Android MDT says:

    Thank you for this article. That’s all I can say. You most definitely have made this blog into something special. You clearly know what you are doing, you’ve covered so many bases.Thanks!

Leave a Reply