This research paper examines the architectural designs and performance tradeoffs of various Cloud Database Management Systems (DBMS). The authors evaluate popular platforms such as Redshift, Vertica, Presto, and Athena using a standardized business analytics workload. A primary focus is the shift from traditional shared-nothing architectures to cloud-native shared-disk models that separate compute power from remote storage. The study analyzes critical factors including initialization times, scalability, and data compatibility across different cloud storage formats. Furthermore, it provides a detailed cost-benefit analysis comparing self-managed instances against serverless DBMS-as-a-service offerings. Ultimately, the findings help organizations choose the right configuration by highlighting how caching, storage types, and licensing impact overall efficiency.

This research explores how B+-tree indexing can be optimized for modern storage hardware that features built-in transparent compression. Traditionally, LSM-trees have been preferred for their superior write amplification and storage efficiency, but the authors argue that hardware-based compression allows for a re-evaluation of this hierarchy. By utilizing sparse data structures that do not consume extra physical space on these specialized drives, the paper proposes three techniques: deterministic page shadowing, localized page modification logging, and sparse redo logging. These methods leverage the decoupling of logical and physical storage to significantly lower the volume of data written to flash memory. Experimental results demonstrate that these optimizations can reduce write amplification by over 10x, effectively closing the performance gap between B+-trees and LSM-trees. Ultimately, the study suggests that advances in computational storage provide a new design space for innovating traditional database management systems.

While Raft is a popular default for distributed consensus, it is often less efficient than Quorums or Reconfiguration. Paxos and Raft mainly differ in leader election. Choosing the right algorithm based on storage efficiency and availability needs can optimize system performance.

Modern NUMA systems impact DBMS performance due to varying memory latency. This research shows that partitioning data and co-locating threads with local memory improves query throughput by up to 75%. Such strategies optimize LLC hit rates and reduce inter-socket communication.

Research reveals that Linux file systems (ext4, XFS, Btrfs) and applications (PostgreSQL, Redis) handle fsync failures inconsistently. File systems often mark pages clean despite write errors, causing data loss or corruption. Applications lack sufficient recovery strategies.

Amazon Aurora Multi-Master enables simultaneous write operations across multiple nodes for increased throughput and continuous availability. Using a decoupled architecture, it employs optimistic conflict detection at the storage layer to ensure data integrity without global locking.

Consensus vs SMR

Monarch is Google’s globally distributed, in-memory time series database designed for planet-scale monitoring. It features a regionalized architecture for high availability and a schematized relational model for complex queries. Key optimizations include field hints and query pushdown.

Consistent replication follows a spectrum: Quorums (failure masking) offer high availability but low efficiency. Reconfiguration (failure detection) maximizes resources but risks downtime. Leaders (hybrid) like Raft balance both, using leases and quorums to manage failures.

Google File System