Amazon DynamoDB vs RocksDB

Amazon DynamoDB is a serverless, managed NoSQL database for applications that need single-digit millisecond performance at any scale, according to AWS. Users do not provision or patch database servers, and tables can scale automatically, including to zero in on-demand mode. Primary keys and global or local secondary indexes provide ways to query data. Global tables support multi-Region, multi-active operation; AWS says they are designed for up to 99.999% availability. The database also supports native vector search for similarity searches and agent memory use cases. Data is encrypted in transit and at rest by default, and IAM policies can restrict access to specific items and attributes. Point-in-time recovery provides continuous, per-second backups and restoration within a configured period of 1 to 35 days. A free tier is listed with 25 WCUs, 25 RCUs, and 25 GB of storage per Region and payer account, among other limits. On-demand capacity is billed per request consumed; provisioned capacity is billed for specified capacity even if usage is lower.

RocksDB is an embeddable, persistent key-value store for applications that need fast storage. It stores keys and values as arbitrary byte arrays, with ordering determined by a comparator chosen by the user. Its log-structured merge design allows configuration of tradeoffs among write, read, and space amplification, and the project describes it as optimized for low-latency storage on flash drives and high-speed disks. The C++ library supports opening and closing databases, reads and writes, merging, and compaction filters. Described uses include database storage engines, application data caches, and embedded workloads. Multi-threaded compactions support databases holding multiple terabytes; the project describes an audience of application servers storing up to a few terabytes on local or remote systems. RocksDB also has a Java wrapper, RocksJava, with artifacts published through Maven. It is free and open source under either GPLv2 or Apache 2.0, at the user's choice. Building requires GCC 11 or later or Clang 10 or later for C++20 support; values must be smaller than 4 GB.