Original reporting by TechRepublic. Quotes and technical details are attributed to Pete Stevens, director at Mythic Beasts.
The Experiment: Serving Millions of Pages from $1,000 of Hardware
When the Raspberry Pi 4 launched in June 2019, the official raspberrypi.org website experienced its highest-ever traffic day. What made the event unusual from a DevOps perspective was that a significant portion of that traffic was served by an 18-board Raspberry Pi 4 cluster — the very hardware being announced — built and operated by Mythic Beasts, the ISP responsible for hosting the site.
The cluster cost less than $1,000 in hardware and handled the record demand without issue. According to Pete Stevens, director at Mythic Beasts, the team served several million pages from the Pi cluster on launch day alone.
“In hindsight we over-specified our cluster — we only really needed about four Pi 4s to do all the WordPress pages (plus two for static content and two for memcache).”
Pete Stevens, Director, Mythic Beasts
Cluster Specifications
The 18-node Pi 4 cluster delivered the following aggregate resources:
- Compute: 72 cores (4 cores per board × 18 boards)
- RAM: 72 GB DDR4 (4 GB per board × 18 boards, using the 4 GB Pi 4 model priced at $55 each)
- Power draw: Under 100 W for the entire cluster
- Physical footprint: Approximately ½U of rack space
Those density and power figures are striking when compared to traditional x86 rack servers. Mythic Beasts noted bluntly that “server manufacturers should start getting scared.”
Why the Pi 4 Is a Step Change for Server Workloads
Earlier Raspberry Pi boards were largely dismissed as serious server candidates due to shared USB and Ethernet buses, limited RAM, and weak single-core performance. The Pi 4 addresses several of those bottlenecks:
- True Gigabit Ethernet — no longer sharing bandwidth with the USB bus
- Up to 4 GB LPDDR4 RAM — enough headroom for WordPress, memcache, and static serving roles
- Roughly 2.5× the CPU performance of the Raspberry Pi 3 B+ (as characterised by Mythic Beasts; independent benchmarks may vary)
The Economics: ARM Cores vs. x86 Cores at Scale
Stevens made a pointed cost comparison for parallelisable workloads such as microservices:
“I can buy a 48-port switch ($750), 48 Pi 4s (48 × $55) and 48 PoE HATs (48 × $30) and for around $5,000 you’ve got 192 ARM cores. In traditional server land you’ll be lucky to get 32 Intel or AMD cores for the same money. So if you’re running microservices you might get significantly better value out of the Raspberry Pi.”
Pete Stevens, Director, Mythic Beasts
The argument is not that a Pi 4 is a drop-in replacement for a high-end server. Rather, it is that certain classes of workload — those that scale horizontally and tolerate lower per-core throughput — can be run on Pi hardware for a fraction of the cost. Stevens frames the question this way:
“The question isn’t ‘Is my Pi 4 a drop-in replacement for my $5,000 servers?’ It’s ‘What applications that currently use $5,000 servers can easily be modified to run on less than $500 of Pi 4 hardware?'”
Pete Stevens, Director, Mythic Beasts
What Was — and Was Not — Running on the Pi Cluster
It is important to understand the scope of the deployment accurately. Not every component of raspberrypi.org ran on the Pi 4 cluster at launch:
- Running on Pi 4: WordPress page serving, static content delivery, memcache
- Still on non-Pi hardware: Front-end load balancers, download servers, apt servers, and databases
- CDN layer: Cloudflare was used to accelerate content delivery (a Cloudflare outage on launch day — which also affected Amazon and Facebook — caused a period of downtime unrelated to the Pi cluster itself)
Current Limitations and the Roadmap
Stevens identified two concrete blockers that prevented a fuller migration at the time of the Pi 4 launch:
1. No Network Boot Support (Yet)
At launch, the Pi 4 lacked support for booting over network-attached storage. This was listed as a planned firmware addition. Without it, Mythic Beasts was reluctant to move the site’s databases onto the cluster, since that role demands “mirrored storage on enterprise SSDs with high write reliability and long write lifetimes” — not SD cards.
2. Network Uplink Constraints
The existing PoE switch infrastructure was designed around the Pi 3, which had a shared USB/Ethernet bus limited to well under 1 Gbps. The Pi 4’s true Gigabit Ethernet was faster than the switches’ uplinks, making the Pi 4 unsuitable as a high-throughput content server until the network was redesigned and upgraded.
“They’re just too fast for our Pi 3 cloud design.”
Pete Stevens, Director, Mythic Beasts
The majority of raspberrypi.org’s traffic comes from OS image downloads, which were still served from Intel hardware. Mythic Beasts indicated that moving this workload to Pi 4 was feasible once the network infrastructure was upgraded.
The Broader Argument: “Good Enough” Technology Wins
Stevens drew a historical parallel to explain why ARM-based commodity hardware may eventually challenge x86 dominance in the server market, despite previous ARM server attempts failing to gain significant share:
“History repeats. Microsoft and Linux won over DEC, Sun, IBM, HP, SGI by being cheap and good enough and steadily getting better.”
Pete Stevens, Director, Mythic Beasts
The pattern is familiar in infrastructure: workloads migrate to cheaper, lower-power platforms as those platforms cross a “good enough” threshold and the ecosystem matures around them. The Pi 4 experiment suggests that threshold may already have been crossed for horizontally scalable web workloads.
Key Takeaways for DevOps Engineers
- Parallelisable workloads (microservices, static serving, caching layers) are the strongest candidates for Pi-based clusters.
- Total cost of ownership calculations should include PoE HATs, switches, and storage — not just the board price.
- Network infrastructure may be the practical bottleneck before compute when deploying Pi 4 at scale.
- Persistent storage strategy matters: SD cards are not suitable for database workloads; network boot and enterprise SSDs are required for production database hosting.
- The Pi 4’s true Gigabit Ethernet is the single most important improvement for server use cases compared to earlier models.
Source: TechRepublic reporting on Mythic Beasts’ deployment, June 2019. Original quotes from Pete Stevens, Mythic Beasts director.