The Hardware Reality: Data Center Server Types
Why this matters: every estimate needs a reference point. "How many servers?" is meaningless until you know what a server is — and the three tiers in a typical design have profoundly different resource profiles.
Key takeaway
Data centers use commodity hardware to scale cost-effectively. The three common server roles — web, application, and storage — need very different amounts of processor, memory, and disk, which is exactly why they are separated in the first place.
The three roles
Web servers
Web servers are the first point of contact after load balancers and typically handle API calls. They are deliberately decoupled from application servers for scalability — so each tier can be sized and scaled to its own bottleneck.
Their memory and storage requirements are often moderate, but they need strong processing power to handle connection setup, TLS, parsing, and routing. Facebook has used web servers with 32 GB RAM and 500 GB storage.
Application servers
Application servers execute business logic and generate dynamic content. They often require significant computational and storage resources. Facebook has deployed application servers with 256 GB RAM and 6.5 TB of hybrid storage (flash and rotating disk).
Storage servers
As data grows, services use specialized storage units. YouTube, for example, uses several distinct systems for distinct jobs:
| Storage system | What it holds | Why that system |
|---|---|---|
| Blob storage | Encoded videos | Large immutable objects, streamed sequentially |
| Temporary processing queue storage | Daily uploads pending processing | Short-lived staging area between upload and encode |
| Bigtable | Video thumbnails | Enormous numbers of small objects, random access |
| RDBMS | Metadata — comments, likes, user channels | Relational queries and transactional integrity |
Other systems, such as Hadoop's HDFS, serve analytics. Storage servers manage both structured (SQL) and unstructured (NoSQL) data.
Returning to Facebook: they have used servers with storage capacity up to 120 TB. Given the number of servers in use, this lets them store exabytes of data. But note that the RAM in these storage servers is often only 32 GB — they are built to hold data, not to compute over it.
The resource profiles, side by side
The three roles differ sharply in what they consume. Approximately, on a low/medium/high scale:
| Server type | Processor | RAM | Hard drive |
|---|---|---|---|
| Web server | High | Medium | Medium |
| Application server | High | High | High |
| Storage server | Low | Low | High |
Read the storage server row against the Facebook numbers: 120 TB of disk paired with 32 GB of RAM. That is a ratio of roughly 4,000:1, and it tells you exactly what the machine is for. Trying to run computation on it would be a category error.
The reference server
Every calculation in this chapter needs a grounding point. This is the typical server we will estimate against:
| Component | Specification |
|---|---|
| Processor | Intel Xeon (Sapphire Rapids 8488C) |
| Number of cores | 64 cores |
| RAM | 256 GB |
| Cache (L3) | 112.5 MB |
| Storage capacity | 16 TB |
Memorize the shape rather than the model number: 64 cores, 256 GB RAM, ~16 TB disk. Those three figures drive nearly every estimate that follows — the core count becomes requests per second, the RAM becomes cache capacity, and the disk becomes how many machines a dataset needs.
Key takeaway
Pick one reference machine and reuse it everywhere. Consistency matters far more than accuracy: an estimate built on one clearly-stated server spec is auditable and correctable, while one built on shifting assumptions cannot be checked at all.
Interview signal by level
| Level | What a strong answer sounds like |
|---|---|
| L4 | Draws "web server" and "database" without a sense of their capacity. |
| L5 | Grounds the design: "assume a 64-core box with 256 GB RAM — that's what I'm sizing against." |
| Staff+ | Matches hardware to role: "storage nodes are disk-heavy and RAM-light, so I won't co-locate compute there. The tiers are separated precisely so each scales on its own bottleneck." |
Next: the numbers that make estimation possible.