Leaky Abstraction
A software engineering concept stating that no abstraction layer can fully hide its underlying implementation details, and that underlying complexity will always leak through to the upper layer in some way.
Core Facts
Timeline (last 90 days)
Typical scenarios of abstraction leakage include distributed training getting stuck at all-reduce, mixed precision causing gradient overflow, and custom operators not being supported.
The core idea of the Law of Leaky Abstractions is that all non-trivial abstractions are leaky, and underlying details will eventually surface, forcing developers to understand them.
Unified computing abstractions may introduce abstraction leakage problems, where the complexity of the underlying implementation penetrates the abstraction layer and is exposed to users
All Facts (3)
The core idea of the Law of Leaky Abstractions is that all non-trivial abstractions are leaky, and underlying details will eventually surface, forcing developers to understand them.
65%UnverifiedTypical scenarios of abstraction leakage include distributed training getting stuck at all-reduce, mixed precision causing gradient overflow, and custom operators not being supported.
50%UnverifiedUnified computing abstractions may introduce abstraction leakage problems, where the complexity of the underlying implementation penetrates the abstraction layer and is exposed to users
50%