Murat Buffalo's blog provides insights into computer science research, machine learning, and artificial intelligence. Readers can explore articles covering topics such as algorithm design, data mining, and computational biology. Additionally, they can learn about machine learning algorithms, deep learning techniques, and applications of AI in various domains.

Metadata

Token bucket algorithms help prevent retry storms in distributed systems, but a subtle concurrency bug can arise when implementing them in sequential client drivers. Using PlusCal and TLA+, this post models the problem: a sequential thread that awaits token availability creates head-of-line blocking, permanently deadlocking the client. The fix is fail-fast behavior — drop the retry instead of waiting. The post also contrasts PlusCal's sequential model with pure TLA+'s event-driven state machine approach, which naturally avoids HOL blocking, and highlights the impedance mismatch between formal specs and real-world sequential implementations.

Modeling Token Buckets in PlusCal and TLA+