数据库选型参考

MySQL Server (open-source relational database under Oracle)

The "default choice" among open-source relational databases — a well-rounded machine for general OLTP, whose extension ecosystem is its deepest moat; the price is internalizing PG's mechanism taxes (VACUUM, the connection model, major-version upgrades) as your team's operational capability.

The de facto standard of enterprise relational databases — and its "tax": you pay for four decades of correctness engineering, the PL/SQL installed base, and the MAA high-availability stack; the price is the industry's most complex licensing metering system and the deepest vendor lock-in.

Microsoft's flagship commercial relational database — the de facto standard OLTP database for the Windows/.NET ecosystem; known for T-SQL, the SSMS toolchain, Always On high availability, and a defense-in-depth security stack. Licensing complexity and per-core cost are questions any evaluation must answer head-on.

The legitimate fork of MySQL (led by founder Monty Widenius). In the early 2010s it was "a drop-in replacement for MySQL"; today it has substantively diverged from Oracle MySQL into two independently evolving products. Generous community edition; the company has just gone through privatization turbulence.

A natively distributed relational database developed in-house by Ant Group, built for financial-grade core systems — one of the top domestic choices for "qu-O" (de-Oracle, replacing Oracle) scenarios.

PingCAP's open-source distributed NewSQL database — the "cloud-native MySQL scaling solution": it gives MySQL-based teams horizontal scaling and real-time HTAP without sharding or building ETL pipelines, but you pay for it in fixed architectural cost and operational complexity.

EDB Postgres Advanced Server is a commercial distribution that bundles the PostgreSQL kernel, an Oracle-migration compatibility layer, and commercial support / HA / security tooling into one enterprise platform. Its value is not in replacing community PostgreSQL, but in **reducing organizational risk for large Oracle migrations and for regulated enterprises standardizing on Postgres**.

A Spanner-style strongly consistent distributed SQL database from Cockroach Labs: Serializable by default, multi-active access across regions, PostgreSQL wire protocol — but "multi-active access" does not mean "cross-ocean writes complete locally," and "PG compatible" does not mean "PostgreSQL."

A distributed SQL database from US-based Yugabyte, Inc.: it lifts PostgreSQL's query layer (the C code) directly onto a self-developed distributed DocDB storage layer (per-tablet Raft + RocksDB). "PG dialect, horizontal scale-out, and multi-region" are its three calling cards; since 2019 its core has been 100% Apache 2.0, forming a hard licensing watershed against CockroachDB, which took its new source private in 2026.

AWS's cloud-native relational database: compute-storage separation, a 6-copy quorum storage layer, fully managed with MySQL / PostgreSQL dual-protocol compatibility — the default answer for teams on AWS that "don't want to operate a database".

Google Cloud's PostgreSQL-compatible cloud-native database: disaggregated compute and storage + dual row-store/column-store engines, positioned as "Aurora on GCP, but PostgreSQL-only with HTAP"; also available as a downloadable AlloyDB Omni edition that runs in any environment.

Alibaba Cloud's self-developed cloud-native relational database, a managed cloud offering on the same track as AWS Aurora: complete compute-storage separation (shared storage, single-writer/multi-reader), with three engines (MySQL / PostgreSQL / Oracle-compatible) sharing the same PolarStore distributed storage foundation; public beta in 2017 — one of the earliest large-scale commercial practices of "cloud-native compute-storage separation" among Chinese cloud vendors.

Tencent Cloud's enterprise-grade database **product family**: it started as financial-grade distributed OLTP (TDSQL for MySQL) and grew into cloud-native Serverless (TDSQL-C), PG/Oracle dual-mode compatibility (TDSQL for PostgreSQL), and HTAP plus a new multi-model engine (TDSQL Boundless) — the complete answer to "which database do I use on Tencent Cloud", except every answer binds you deeply to Tencent Cloud.

Google Cloud's globally distributed relational database, which uses TrueTime (GPS + atomic-clock timekeeping) to make "cross-continent linear consistency" an engineering reality — it is one of the very few systems to bring external consistency (strict serializability) to global OLTP scale, at the cost of GCP lock-in, a high entry floor, and the design discipline of "modeling is sharding".

The de facto standard for in-memory data structure stores: trading "all-in-memory + single-threaded event loop + rich data structures" for sub-millisecond latency; after the 2024 licensing upheaval it split into two lines — Redis (Redis Inc., triple-licensed) and Valkey (Linux Foundation, BSD) — protocol-compatible but with new features drifting apart.

CNCF-graduated distributed KV store with Raft strong consistency, aimed at metadata / coordination / configuration scenarios — the metadata foundation of Kubernetes. "Boring reliability" itself: keep your ambitions within metadata and it almost never takes the blame; point your ambitions at the data plane and every one of its design constraints becomes a title in your postmortem.

AWS's serverless KV/document database: sharded by partition-key hash, single-digit-millisecond latency independent of table size, multi-active across regions — but "no ops" comes at the cost of turning the DBA's job into capacity modeling and bill engineering, and designing your access patterns wrong costs more than picking the wrong database.

The original wide-column distributed NoSQL and its C++ rewrite: storage built for "write-heavy, always-on, horizontally scalable" workloads, with tunable consistency on an eventually consistent base. Cassandra stands for ecosystem maturity and license neutrality; ScyllaDB stands for per-node performance and tail-latency control — at the cost of a source-available license since 2025 and a smaller community.

The de facto standard for document databases: BSON documents + flexible schema + native sharding, taking "time-to-launch" to the extreme — but "no schema" does not mean "no data governance," and the SSPL license means it is long past being an open-source database in the traditional sense.

A natively distributed relational database developed in-house by Ant Group, built for financial-grade core systems — one of the top domestic choices for "qu-O" (de-Oracle, replacing Oracle) scenarios.

An open-source columnar OLAP database for high-throughput append writes and massive scan-aggregation — "a real-time analytics engine with extreme single-node efficiency": columnar storage, vectorized execution and MergeTree at its core, delivering scan throughput on one box that row-store warehouses need clusters for; the price is almost no general-purpose ACID transactions, architecturally expensive high-frequency updates/deletes, and distributed replication plus DDL operations you must design yourself.

Apache Doris is an Apache top-level project (originating from Baidu Palo), an MPP real-time analytical database: columnar storage plus a fully vectorized execution engine, accessed via the MySQL protocol. It targets "unified serving of high-concurrency point queries and complex OLAP queries" and is a mainstream choice for real-time data warehousing and unified analytics.

A blazing-fast MPP analytical database with roots in the Baidu Palo / Doris codebase — built for "real-time reporting over many joined tables plus lakehouse in one engine": a fully vectorized execution engine with a Cascades-style CBO, a primary-key model for real-time upserts, and shared-data storage for elasticity; the price is front-loading constraints like FE-centralized metadata, tablet planning, and primary-key memory — decisions made on day one of table design.

The "analytical SQLite" — an in-process, single-node columnar OLAP engine: `pip install` and go, vectorized execution saturating all cores, SQL directly against Parquet/CSV/S3. It devours the middle ground between "not enough for a warehouse, too slow in pandas" — at the cost of a single-writer lock, a single-node ceiling, and zero server-side capabilities (auth, audit, multi-user).

A cloud-native Lakehouse platform founded by the original Apache Spark team — "data-lake scale + warehouse reliability + a native foundation for AI": Delta Lake (ACID transaction log) unifies storage, Unity Catalog unifies governance, and the Photon vectorized engine accelerates queries; one platform covers data engineering, BI, machine learning and AI agents. The price is pure cloud consumption billing (DBUs + cloud-resource double invoice), the Spark tuning burden, and a "data is open, the platform is proprietary" lock-in structure.

The SaaS benchmark for cloud-native warehousing — "an elastic data cloud with storage fully separated from compute": a three-layer architecture (storage/compute/cloud services) lets N independent warehouses read the same data with zero contention, while Time Travel + zero-copy cloning turn "oops recovery" into a daily operation. The price is pure SaaS (no self-hosting), cost unpredictability from consumption-based credits, and real migration cost out of a proprietary storage format.

AWS's veteran cloud data warehouse — born from the ParAccel columnar MPP engine, disaggregated storage and compute with RA3 in 2019, Serverless GA in 2022: Spectrum queries S3 data lakes directly, Zero-ETL streams from Aurora/DynamoDB, data sharing serves live data across clusters. The price is deep AWS lock-in, the traditional MPP physical-design mindset (distkey/sortkey/vacuum), and a Serverless RPU bill that still needs governance.

Google's serverless warehouse benchmark — Dremel-blooded storage-compute separation, no nodes, no clusters, queries auto-allocated slots: two billing modes (per scanned TiB on-demand, or per slot-hour under Editions), Omni ships the engine to where AWS/Azure data lives. The price is deep GCP lock-in, "one bad SQL burns the budget" under scan-based billing, and slot governance as mandatory homework.

A natively distributed relational database developed in-house by Ant Group, built for financial-grade core systems — one of the top domestic choices for "qu-O" (de-Oracle, replacing Oracle) scenarios.

Zilliz's open-source distributed vector database, built for billion-scale vector search: its index-type matrix (HNSW / IVF / DiskANN / sparse / BM25) and storage-compute separation are the real deal — but the complexity is split across coordination, message queues, object storage, segments, and index building. If your data volume hasn't pushed past pgvector's limits, those capabilities are just extra ops tax.

Weaviate is an open-source, AI-native vector database: it folds vector retrieval, BM25 hybrid search, embedding/rerank/generation model modules, and multi-tenant lifecycle management into one system. Its sweet spot is RAG and semantic-search platforms — not a transactional system of record.

An open-source vector database written in Rust by Qdrant Solutions (Berlin) — built for AI applications (RAG, semantic search, recommendations, agent memory) that need "low-latency approximate nearest-neighbor search with rich metadata filtering": its core strengths are filterable HNSW, multi-vector/sparse vectors with server-side hybrid retrieval fusion, a tunable quantization matrix, and a payload-driven multitenancy model; but it is a **derived index at the retrieval layer**, not a system of record — no multi-record ACID transactions, no continuous PITR, and a two-tier distributed consistency model ("Raft for metadata, tunable acknowledgements for point data") that you must internalize to use it correctly.