arXiv subjects

Remzi Arpaci-Dusseau

Publications and source records attributed to Remzi Arpaci-Dusseau.

6 recordsLinked to original sources

arXiv2026

InfraBench: Evaluating Infrastructure Agents Across Layers, Lifecycle, and Risk

Managing modern computing infrastructure has become a steadily harder problem due to the ever-increasing complexity. Recent advances in AI agents create a timely opportunity to automate infrastructure management tasks, but it remains unclear how well such agents can handle real-world infrastructure complexity. We present InfraBench, a benchmark suite for evaluating AI agents on realistic infrastructure tasks across the full system stack and full operational lifecycle with fine-grained risk assessment. Experiments with 15 agent-model configurations show that even the strongest agent cannot secure a full score across all tasks. Mean effective scores range from roughly 40% to 88% (with per-configuration standard errors of 6-12 points), repeating every task three times reveals that top configurations still pass only a fraction of their attempts, and per-check scoring exposes a general failure pattern: agents may routinely satisfy short-term objectives while leaving non-durable changes, broken distributed invariants, unsafe side effects, and uncleaned state behind. INFRABENCH, including its live leaderboard, tasks, and evaluation harness, is publicly available at infraben.ch.

cs.AI↗

arXiv2025

Crossword: Adaptive Consensus for Dynamic Data-Heavy Workloads

We present Crossword, a flexible consensus protocol for dynamic data-heavy workloads, a rising challenge in the cloud where replication payload sizes span a wide spectrum and introduce sporadic bandwidth stress. Crossword applies per-instance erasure coding and distributes coded shards intelligently to reduce critical-path data transfer significantly when desirable. Unlike previous approaches that statically assign shards to servers, Crossword enables an adaptive tradeoff between the assignment of shards and quorum size in reaction to dynamic workloads and network conditions, while always retaining the availability guarantee of classic protocols. Crossword handles leader failover gracefully by employing a lazy follower gossiping mechanism that incurs minimal impact on critical-path performance. We implement Crossword (along with relevant protocols) in Gazette, a distributed, replicated, and protocol-generic key-value store written in async Rust. We evaluate Crossword comprehensively to show that it matches the best performance among previous protocols (MultiPaxos, Raft, RSPaxos, and CRaft) in static scenarios, and outperforms them by up to 2.3x under dynamic workloads and network conditions. Our integration of Crossword with CockroachDB brings 1.32x higher aggregate throughput to TPC-C under 5-way replication. We will open-source Gazette upon publication.

cs.DC↗

arXiv2025

Bodega: Serving Linearizable Reads Locally from Anywhere at Anytime via Roster Leases

We present Bodega, the first consensus protocol that serves linearizable reads locally from any desired node, regardless of interfering writes. Bodega achieves this via a novel roster leases algorithm that safeguards the roster, a new notion of cluster metadata. The roster is a generalization of leadership; it tracks arbitrary subsets of replicas as responder nodes for local reads. A consistent agreement on the roster is established through roster leases, an all-to-all leasing mechanism that generalizes existing all-to-one leasing approaches (Leader Leases, Quorum Leases), unlocking a new point in the protocol design space. Bodega further employs optimistic holding and early accept notifications to minimize interruption from interfering writes, and incorporates smart roster coverage and lightweight heartbeats to maximize practicality. Bodega is a non-intrusive extension to classic consensus; it imposes no special requirements on writes other than a responder-covering quorum. We implement Bodega and related works in Vineyard, a protocol-generic replicated key-value store written in async Rust. We compare it to previous protocols (Leader Leases, EPaxos, PQR, and Quorum Leases) and two production coordination services (etcd and ZooKeeper). Bodega speeds up average client read requests by 5.6x-13.1x on real WAN clusters versus previous approaches under moderate write interference, delivers comparable write performance, supports fast proactive roster changes as well as fault tolerance via leases, and closely matches the performance of sequentially-consistent etcd and ZooKeeper deployments across all YCSB workloads. We will open-source Vineyard upon publication.

cs.DC↗

arXiv2025

A Unified, Practical, and Understandable Model of Non-transactional Consistency Levels in Distributed Replication

We present a practical model of non-transactional consistency levels in the context of distributed data replication. Unlike prior work, our simple Shared Object Pool (SOP) model defines common consistency levels in a unified framework centered around the single concept of ordering. This naturally reflects modern cloud object storage services and is thus easy to understand. We show that a consistency level can be intuitively defined by specifying two types of constraints on the validity of orderings allowed by the level: convergence, which bounds the lineage shape of the ordering, and relationship, which bounds the relative positions between operations. We give examples of representative protocols and systems, and discuss their availability upper bound. To further demonstrate the expressiveness and practical relevance of our model, we use it to implement a Jepsen-integrated consistency checker for the four most common levels (linearizable, sequential, causal+, and eventual); the checker analyzes consistency conformity for small-scale histories of real system runs (etcd, ZooKeeper, and RabbitMQ).

cs.DC↗

arXiv2025

Zerrow: True Zero-Copy Arrow Pipelines in Bauplan

Bauplan is a FaaS-based lakehouse specifically built for data pipelines: its execution engine uses Apache Arrow for data passing between the nodes in the DAG. While Arrow is known as the "zero copy format", in practice, limited Linux kernel support for shared memory makes it difficult to avoid copying entirely. In this work, we introduce several new techniques to eliminate nearly all copying from pipelines: in particular, we implement a new kernel module that performs de-anonymization, thus eliminating a copy to intermediate data. We conclude by sharing our preliminary evaluation on different workloads types, as well as discussing our plan for future improvements.

cs.OS↗

arXiv2024

Foreactor: Exploiting Storage I/O Parallelism with Explicit Speculation

We introduce explicit speculation, a variant of I/O speculation technique where I/O system calls can be parallelized under the guidance of explicit application code knowledge. We propose a formal abstraction -- the foreaction graph -- which describes the exact pattern of I/O system calls in an application function as well as any necessary computation associated to produce their argument values. I/O system calls can be issued ahead of time if the graph says it is safe and beneficial to do so. With explicit speculation, serial applications can exploit storage I/O parallelism without involving expensive prediction or checkpointing mechanisms. Based on explicit speculation, we implement Foreactor, a library framework that allows application developers to concretize foreaction graphs and enable concurrent I/O with little or no modification to application source code. Experimental results show that Foreactor is able to improve the performance of both synthetic benchmarks and real applications by significant amounts (29%-50%).

cs.OS↗