Reading Stacks Season 2: Distributed Paper Stacks

In this episode of Paper Stacks, Can Toraman will be discussing Achieving Task-Based Scheduling in Low-Latency Distributed Priority Queues.

In modern distributed systems, managing the queue efficiently is crucial for performance and responsiveness. Especially in multi-tenant environments, prioritizing different tasks (based on customer, task type, etc) is crucial. Yet, this area is still not highly explored, as many companies solve it by using different queues. Well, are more optimal ways, and how are they achievable in a distributed setting?

Distributed priority queues are still not highly explored in the academic environment, so will base our discussions on different theoretical and practical perspectives used in the industry:

• Timestone, Netflix’s distributed queue used in its encoding service
• ⁠B4, Google’s Wide Area Network to support requests from both GCP and Google applications.
• ⁠Facebook’s FOQS to support horizontal scaling and multi-tenant environments
• ⁠QPID: A Distributed Priority Queue with Item Locality

Links:

• Timestone: https://netflixtechblog.com/timestone-netflixs-high-throughput-low-latency-priority-queueing-system-with-built-in-support-1abf249ba95f
• B4: https://research.google/pubs/b4-and-after-managing-hierarchy-partitioning-and-asymmetry-for-availability-and-scale-in-googles-software-defined-wan/
• FOQS: https://engineering.fb.com/2021/02/22/production-engineering/foqs-scaling-a-distributed-priority-queue/
• QPID: https://ieeexplore.ieee.org/document/4725152

In this episode of Paper Stacks, Ozan Sazak will explore the internals of Amazon Dynamo, a highly available and scalable key-value store.

Dynamo, developed for Amazon’s critical services, addressed the internal needs of Amazon using techniques such as consistent hashing, quorum-like techniques, and eventual consistency.

We’ll dive into the 2007 Dynamo paper, discuss the internals of the Dynamo, and the design decisions between the parts such as partitioning and replication.

In this episode of Paper Stacks, Alperen Keles and Ozan Akin will be discussing Real-Time Distributed State Sharing for Collaborative Applications using Conflict-Free Replicated Data Types(CRDTs).

Sharing state between multiple servers and clients is extremely common today. When multiple users make concurrent updates on a shared data structure, the applications need conflict resolution strategies to ensure all users agree on the final state.

CRDTs are one such mechanism for handling distributed state; the differentiating factors for CRDTs are (1) they don’t need a central server, (2) they are highly fault-tolerant, allowing users to work offline, and sync their work at arbitrary intervals.

Within the episode, we will base our discussions on the 2011 paper by Marc Shapiro, discuss the usage of CRDTs in collaborative text editing in Zed Code Editor, collaborative canvases such as Excalidraw and Figma, and provide live demos on the internal workings of CRDTs.

In this episode of Paper Stacks, Yigit Varli will be discussing the challenges of optimizing tail latencies in large-scale distributed systems.

Distributed systems tend to serve a diverse set of users at geographically varying locations, with heterogeneous loads and requirements. Classically, such systems have optimized for “average latency”, which resulted in very high latency for a small subset of users, the so-called tail.

In this episode, we will dive into ‘The Tail at Scale’ by Jeffrey Dean and Luiz André Barroso from Google, where we will explore system designs and techniques to handle high loads, ensuring smoother operation and improved user experience.