Bending Spoons with Flexible Fastly: 50% latency reduction plus significant egress cost savings

Bending Spoons has a history of taking the impossible in its stride — developing the tech to power a portfolio of digital products delighting nearly a billion users around the globe.

And with products as diverse as a top ranked personal productivity app Evernote and AI-powered photo enhancement and generation tool Remini, it’s vital to find the right edge cloud platform partner to ensure each product looks fantastic, is easy to use, and rock-solid under the hood.

With an ever-increasing user-base and a team managing a growing number of products, it soon became a priority to find a partner expert in networking who could offer competitive terms while providing a reliable and performant origin-independent service for critical workloads.

Acquiring and transforming a tech product in a broad portfolio is like finally getting a coveted car in your workshop — you don’t know the exact scale of the job until you pop the hood and cast an expert eye over the engine. As Bending Spoons Platform Engineer Davide Pedranz explains: “Let’s take Cabinet, for example. It was small and we only acquired the assets, so we managed to rebuild it from zero to close to where the app is now in a couple of months. However, Evernote was different because we were acquiring the whole company, the product was so much more complex, and it had been around for 15 years. There were around 10 million lines of code, so it’s not something you can casually read and rewrite overnight.”

From caching and load balancing technologies to enterprise support, Bending Spoons wanted to invest in a platform powerful enough to cover all existing use cases AND versatile enough to support ambitious growth through future acquisitions.

Stacked challenges

Regularly acquiring products means your tech stack can expand exponentially. So, achieving smooth, streamlined integration is another significant challenge.

Bending Spoons started with Google Cloud Platform (GCP).

Its base stack was Cloud DNS to manage DNS records, Cloud CDN (when required), external HTTP load balancer , managed instance groups (VM) on Compute Engine, and various databases (depending on use case). As well as a caching solution when needed (typically Redis), they also used GCS for storing files, Pub/Sub for async queues, and BigQuery for analytics.

And talented, diligent ‘Spooners’ are key to achieving its ‘impossible mission’. Every team runs its own infrastructure, and each can use an internal library of opinionated infrastructure components (written as Terraform modules). This flexible, agile setup gives teams independence and velocity yet ensures the resulting infrastructures are reasonably similar.

Achieving consistency and integration is challenging and has become more complex with a recent increase in acquisitions. As Davide says: “We try to refactor or rebuild the infrastructure to look similar to our internal standards. For example, after the Evernote acquisition, we rebuilt the entire production (and non-production) infrastructure from scratch in a few months.”

Through the Evernote acquisition, Bending Spoons inherited the Akamai platform. Although it was delivering content, its main use case had been security. However, a new solution was required which combined updating the underlying service and deploying smarter attenuation measures (rate limits and request tarpitting).

After contacting the main market players, Bending Spoon approached Fastly, which could provide more intuitive UI/Terraform modules, much faster deployments of configuration changes, and more powerful primitives — for example, tarpitting was excellent given the backend constraints. “We knew Fastly were the cool kids on the market,” says Davide, “and the CDN platform looked legit, so we tried it, liked it a lot and found commercials that were good for both parties. So, it was like, okay, let’s go. When Akamai found out we wanted to switch, they offered us a price 10% lower. But we didn’t want to play that game. We wanted a great platform, and we found Fastly superior to Akamai in all aspects.”

Fastly tried and tested

Bending Spoons wanted a vendor with crucial criteria including custom configurations, service reliability, time from a configuration change to going live, good support for Terraform, APIs, and documentation.

Once reasonable commercials were secured, Davide’s team tested configurations for the main use cases of services it planned to migrate to Fastly — non-trivial VCL configurations, monitoring with Prometheus, and logging to BigQuery.

The entire process was swift and smooth, with commercial discussions ongoing for 3-4 months, the technical demo taking two weeks, and migrating the initial services even before signing the contract.

Now it’s in place, Fastly CDN has two main use cases: 

For CDN.  Files are often static assets stored on Cloud Storage (GCS). Sometimes, they’re dynamically generated by web applications and cached depending on the cache control headers provided by the backend. Bending Spoons uses aggressive surrogate control policies, relying on Fastly’s powerful cache invalidation capabilities when updating that content.

For egress. Bending Spoons isn’t interested in caching at the edge in this setup. Instead, it deploys a Fastly VCL service to serve traffic from its cloud provider to the internet through Fastly. 

Stunning Remini results 

With Fastly firing on all cylinders, the most striking results in the Bending Spoons portfolio thus far are from Remini — a product which helps users improve their images and videos using proprietary AI models.

After the AI has worked its magic, the client uses a signed URL to download the enhanced images from GCS. And Fastly makes a massive difference at this point. Previously, the image was downloaded directly from GCS but now, the signed URL points to a custom domain managed by Fastly, and Fastly fetches the image from GCS and serves it to the client. Caching is explicitly disabled as the content is private to the user.

Let’s take a look at the results:

Davide explains: 

“The content is served from Google to Fastly in the US (through shielding) then traffic goes to Amsterdam, and from Amsterdam it exits to the public internet towards India. With this chain, I would expect the latency you can introduce going from Amsterdam to India to be significantly higher than what you would have from Google to India on their private network. Instead, I was very happy to see that was not the case due to the Argo Smart Routing and Fastly performance network. So that was one result. After solving India, we deployed it worldwide. And the numbers are also quite interesting – it’s roughly 50% faster than it was before.”

“In the second plot here,” says Davide, “it’s divided by platform and you can see as we roll out it to more and more people from the download step (the one impacted by this migration) the median end-to-end latency to download an image from a GCS bucket decreased by ~50%.”

The exact figure for possible savings will depend on your commercial arrangements with your cloud provider. Fastly and Google partnership offered lower egress fees when sending traffic from GCP to Fastly. Depending on the traffic volume, savings were significant.

Around the next bend

Bending Spoons’ acquisition model continues apace, and WeTransfer is the next product the Spooners will get their hands on. And although it’s too early to say exactly how Fastly might help, the platform is now a firm fixture in the toolkit. 

“On top of being performant, Fastly is also very flexible,” says Davide. “Even if you stick to the old VCL approach, you can do some very fancy things which aren’t available from typical cloud providers. Fastly also supports a lot of logging options with BigQuery, and we can pull metrics from Fastly services into Prometheus for monitoring. I expect we’ll continue to find new use cases for Fastly internally for our platform and for products we acquire or launch in the future.”