The Architecture of Computation: Why Computer Science is a Superset of Data Science

Formal Proposition

Data science models empirical observations $\mathcal{D} = \{(x_i, y_i)\}$. Computer science models the fundamental nature of transformation $\mathcal{M}: \Sigma^* \to \Sigma^*$, decidability, and state under bounded time and space complexity.

In recent years, the colloquial tech lexicon has increasingly conflated Data Science with the broader discipline of Computer Science. With the explosion of parameter-heavy transformer architectures, statistical gradient descent, and high-level Python wrappers, it is easy for practitioners to view computing through the narrow lens of matrix multiplications and empirical error minimization.

However, from first principles, Computer Science is the foundational superset. Data science represents a specialized, applied branch—an intersection of statistics, numerical linear algebra, and empirical approximation—operating within the computational substrate that Computer Science defines, optimizes, and proves.

1. The Substrate of Computation

Every statistical algorithm, neural loss gradient, and distributed data frame depends on a hierarchy of physical and theoretical constraints governed strictly by computer systems engineering:

Information Theory & Complexity: Before one can optimize a model, one is bound by the asymptotic bounds of algorithmic complexity: $\mathcal{O}(n \log n)$, $\mathcal{P} \text{ vs } \mathcal{NP}$, and Shannon entropy limits.
Memory Hierarchy & Cache Locality: A mathematically optimal algorithm that disregards cache lines ($L1/L2/L3$), NUMA memory nodes, and PCIe bandwidth will be orders of magnitude slower than a cache-conscious implementation.
Deterministic State vs Stochastic Inference: While machine learning models produce probabilistic approximations $\mathbb{P}(y \mid x)$, the underlying database engines, transaction managers, and execution pipelines require strict determinism (ACID properties, serializability, and consensus).

2. The Limits of Decidability and Consensus

Where data science asks “What pattern best fits this historical distribution?”, computer science asks “Is this problem computable, and can distributed nodes agree upon it in the presence of arbitrary network partitions?”

               ┌────────────────────────────────────────────────┐
               │              COMPUTER SCIENCE                  │
               │  ┌──────────────────────────────────────────┐  │
               │  │  Theory of Computation & Formal Automata │  │
               │  │  Distributed Systems & Consensus (Raft)  │  │
               │  │  Memory Subsystems & Hardware Arch       │  │
               │  │  Compiler Design & Type Systems          │  │
               │  │                                          │  │
               │  │   ┌───────────────────────────────────┐  │  │
               │  │   │          DATA SCIENCE             │  │  │
               │  │   │  - Statistical Inference          │  │  │
               │  │   │  - Empirical Regression           │  │  │
               │  │   │  - Loss Optimization (SGD)        │  │  │
               │  │   └───────────────────────────────────┘  │  │
               │  └──────────────────────────────────────────┘  │
               └────────────────────────────────────────────────┘

Consider the classical FLP Impossibility Result (Fischer, Lynch, Paterson, 1985):

In an asynchronous network, no deterministic consensus protocol can guarantee both safety and liveness in the presence of even a single unannounced process crash.

No amount of data collection, feature engineering, or statistical sampling can violate the FLP theorem or bypass the CAP theorem. These are immutable structural laws of information flow and state synchronization.

3. Why Grounding Matters in Modern Software

When engineers treat data science as distinct from foundational computer science, architectural failures inevitably follow:

The “Black-Box” Pipeline: Stacking stochastic models on top of un-isolated data stores with zero transactional guarantees.
Algorithmic Inefficiency: Replacing an $\mathcal{O}(1)$ hash trie or bloom filter with an ungrounded multi-billion parameter model that consumes $100\times$ more energy to perform string matching.
Resilience & Fault Tolerance: Failing to design for network partitions, backpressure, and idempotency when distributed nodes experience partial failure.

Conclusion

Data science provides powerful tools for statistical estimation and pattern synthesis. But without the rigorous foundation of Computer Science—finite automata, formal grammars, algorithmic bounds, memory architectures, and distributed consensus—those tools lack a stable foundation.

Understanding computation from the silicon up to the complexity class is what separates ephemeral hype from engineering that endures.