Adopting a Serverless Approach at Bazaarvoice with MongoDB Atlas and AWS Lambda

Facebook ShareLinkedin ShareReddit ShareTwitter Share

I recently had the pleasure of welcoming Ani Hammond, Senior Staff Software Engineer from Bazaarvoice, to the MongoDB World stage. To a completely packed room, Ani chronicled her team’s journey as they replatformed Bazaarvoice’s Curations service from a runaway monolith architecture to a completely serverless architecture backed by MongoDB Atlas.

Bazaarvoice logo

Even if you’ve never heard of Bazaarvoice, it’s almost impossible that you’ve never interacted with their services. To use Ani’s own description, “If you're shopping online and you’re reading a review, it's probably powered by us.”

Bazaarvoice strives to connect brands and retailers with consumers through the gathering, curation, and display of user-generated content—anything from pictures on Instagram to an online product review—during a potential customer’s buying journey.

To give you a sense of the scale of this task, Bazaarvoice clocked over a billion total page views between Thanksgiving Day and Cyber Monday in 2017, peaking at around 6,000 page views per second!

Even if you’ve never heard of Bazaarvoice, it’s almost impossible that you’ve never interacted with their services.

One of the technologies behind this herculean task is the Curations platform. To understand how this platform works, let’s look at an example:

An Instagram user posts a cute photo of their child wearing a particular brand’s rain boots. Using Curations, that brand is watching for specific content that mentions their products, so the social collection service picks up that post and shows it to the client team in the Curations application. The post can then be enriched in various manual and automatic ways. For example, a member of the client team can append metadata describing the product contained in the image or automatic rules can filter content for potentially offensive material. The Curations platform then automates the process of securing the original poster’s permission for the client to use their content. Now, this user-generated content is able to be displayed in real time on the brand’s homepage or product pages to potential customers considering similar products.

In a nutshell, this is what Curations does for hundreds of clients and hundreds of thousands of individual content pieces.

The technology behind Curations was previously a monolithic Python/Django-based stack on Amazon EC2 instances on top of a MySQL datastore deployed via RDS.

The technology behind Curations was previously a monolithic Python/Django-based stack on Amazon EC2 instances on top of a MySQL datastore deployed via RDS.

This platform was effective in allowing Bazaarvoice to scale to hundreds of new clients. However, this architecture did have an Achilles heel: each additional client onboarded to Bazaarvoice’s platform represented an additional Python/Django/MySQL cluster to manage. Not only was this configuration expensive (approximately $60,000/month), the operational overhead generated by each additional cluster made debugging, patching, releases, and general data management an ever-growing challenge. As Ani put it, “Most of our solutions were basically to throw more hardware/money at the problem and have a designated DevOps person to manage these clusters.”

One of the primary factors in selecting MongoDB for the new Curations platform was its support for a variety of different access patterns. For example, the part of the platform responsible for sourcing new social content had to support high write volume whereas the mechanism for displaying the content to consumers is read-intensive with strict availability requirements.

Diving into the specifics of why the Bazaarvoice team opted to move from a MySQL-based stack to one built on MongoDB is a blog post for another day. (Though, if you’d like to see what motivated other teams to do so, I recommend How DevOps, Microservices, and MongoDB are Making HSBC “Simpler, Better, and Faster” and Breuninger delivers omnichannel shopping experience for thousands of daily online users.)

That is to say, the focus of this particular post is the paradigm shift the Curations team made from a linearly-scaling monolith to a completely serverless approach, underpinned by MongoDB Atlas.

The new Curations platform is broken into three distinct services for content collection, enrichment, and display. The collections service is powered by a series of AWS Lambda functions triggered by an Amazon Kinesis stream written in Node.js whereas the enrichment and display services are built on autoscaling AWS Elastic Beanstalk instances. All three services making up the new Curations platform are backed by MongoDB Atlas.

Not only did this approach address the cluster-per-customer challenges of the old system, but the monthly costs were reduced by nearly 90% to approximately $6,500/month. The results are, again, best captured by Ani’s own words:

Massive cost savings, huge performance gains, strong consistency, and a handful of services rather than hundreds of clusters.

MongoDB Atlas was a natural fit in this new serverless paradigm as the team is fully able to focus on developing their product rather than on infrastructure management. In fact, the team had originally opted to manage the MongoDB instances on AWS themselves. After a couple of iterations of manual deployment and management, a desire to gain even more operational efficiency and increased insight into database performance prompted their move to Atlas. According to Ani, the cost of migrating to and leveraging a fully managed service was, "Way cheaper than having dedicated DevOps engineers.” Atlas’ support for direct VPC peering also made the transition to a hosted solution straightforward for the team.

Speaking of DevOps, one of the first operational benefits Ani and her team experienced was the ability to easily optimize their index usage in MongoDB. Previously, their approach to indexing was “build stuff that makes sense at the time and is easy to iterate on.” After getting up and running on Atlas, they were able to use the built-in Performance Advisor to make informed decisions on indexes to add and unused ones to remove. As Ani puts it:

An index killed is as valuable as an index added. This ensures all your indexes to fit into memory and a bad index doesn't push out the good ones.

Ani’s team also used the Atlas Performance Advisor to diagnose and correct inefficient queries. According to her, the built-in tools helped keep the team honest, "[People] say, ‘My database isn't scaling. It's not able to perform complex queries in real time...it doesn't work.’ Fix your code. The hardware is great, the tools are great but they can only carry you so far. I think sometimes we tend to get sloppy with how we write our code because of how cheap and how easy hardware is but we have to write code responsibly too.”

In another incident, a different Atlas feature, the Real Time Performance Panel, was key to identifying an issue with high load times in the display service. Some client’s displays were taking more than 6 seconds to load. (For context, content delivery network provider, Akamai, found that a two-second delay in web page load time can cause bounce rates to double!) High-level metrics in Datadog reported 5+ seconds query response times, while Atlas reported less than 100 ms response times for the same query. The team used both data points to triangulate and soon realized the discrepancy was a result of the time it took for Lambda to connect to MongoDB for each new operation. Switching from standard Lambda functions to a dockerized service ensured each operation could leverage an open connection rather than initiating a “cold start.”

I know a lot of the cool things that Atlas does can be done by hand but unless this is your full-time job, you're just not going to do it and you’re not going to do it as well.

Ani’s team also used the Atlas Performance Advisor to diagnose and correct inefficient queries.

Before wrapping up her presentation, Ani shared an improvement over the old system that the team wasn’t expecting. Using Atlas, they were able to provide the customer support and services teams read-only views into the database. This afforded them deeper insight into the data and allowed them to perform ad-hoc queries directly. The result was a more proactive approach to issue management, leading to an 80% reduction in inbound support tickets.

By re-architecting their Curations platform, Bazaarvoice is well-positioned to bring on hundreds of new clients without a proportional increase in operations work for the team. But once again, Ani summarized it best:

As the old commercial goes… ‘Old platform: $60,000. New platform: $6,000. Getting to focus all of my time on development: priceless.'

Thank you very much to Ani Hammond and the rest of the Curations team at Bazaarvoice for putting together the presentation that inspired this post. Be sure to check out Ani’s full presentation in addition to dozens of other high-quality talks from MongoDB World on our YouTube channel.

If you haven’t tried out MongoDB Atlas for yourself, you can started with a free sandbox cluster.