Databases in 2025: A Year in Review

Original Article ↗ Hacker News Discussion ↗

MCP, LLM Agents, and Database Security

  • Several commenters agree the MCP “max context” philosophy conflicts with least-privilege and revives SQL injection risks, now via LLM hallucinations instead of malicious users.
  • Proposed mitigations:
    • Gateways that track provenance/trust of context entering the LLM and block “lethal trifecta” actions.
    • Giving agents access only to isolated MVCC snapshots or nested transactions so they can’t commit to the real DB (criticized as non-trivial, easy to mis-implement, and requiring new code to be written/tested).
  • Nobody in the thread reports seriously using write-capable DB MCP in production; some restrict agents to read-only views.

EdgeDB/Gel and the Vercel Acquisition

  • The Gel → Vercel line is seen as misleading: users perceive Gel as effectively sunset, with the team moving to other work.
  • Community is organizing a “blessed” fork to keep Gel alive, with a working group and possible final upstream handoff path.

SQLite, DuckDB, and Embedded / Simpler Architectures

  • Multiple comments describe a strong trend toward SQLite (and sometimes DuckDB) for many workloads: single-file DBs, low operational overhead, JSON support, and excellent local performance.
  • DuckDB is praised for analytics: columnar storage, Parquet/S3, WASM, JSON handling, vector search, and the ability to read/write SQLite files.
  • A long subthread explores patterns like:
    • SQLite for OLTP + DuckDB for OLAP, possibly with periodic batch sync and watermark-based freshness.
    • Tension between wanting low-latency analytics vs. the complexity of maintaining both row- and column-oriented views.
  • Debate over SQLite “in production”:
    • Many argue it’s already one of the most widely used production DBs (mobile/desktop, local apps, small web services).
    • Others stress its limitations for heavy multi-writer workloads and note that traditional client/server RDBMS still dominate larger multi-user deployments.
    • People discuss WAL + BEGIN CONCURRENT, edge-sharding (e.g., Cloudflare D1), and caching tradeoffs.

Immutable, Bi-temporal, and Time-Series Databases

  • Some see a blind spot in the retrospective: no real treatment of immutable/bi-temporal systems (Datomic, XTDB) or dedicated time-series DBs.
  • Advocates argue these are particularly valuable in fintech/healthcare for auditability, regulatory compliance, time-travel queries, and fast undo.
  • Others note you can approximate many of these properties in Postgres (range types, bitemporal extensions, audit triggers plus WAL archiving).
  • Time-series-specific innovation is perceived as relatively quiet; ClickHouse and QuestDB are mentioned, but commenters wish for more coverage.

PostgreSQL Dominance, MySQL/MariaDB, and Enterprise Mainstays

  • Some question labeling PostgreSQL “dominant” given metrics that still show larger MySQL/MariaDB installed bases (e.g., via WordPress).
  • Counterpoints:
    • The “dominance” being discussed is about momentum, mindshare, investment, and new-project choices, not raw install count.
    • Rumors of major layoffs in the MySQL team at Oracle and MariaDB Corp’s financial troubles are said to be nudging organizations toward Postgres.
    • Many new systems (e.g., Spanner interfaces, ClickHouse) choose PostgreSQL compatibility now rather than MySQL.
  • Others note that traditional enterprise still runs heavily on Oracle, SQL Server, DB2, or SaaS abstractions, and these are underrepresented because the review focuses on “what’s changing,” not stable incumbents.

Omissions and Niche Systems

  • Multiple commenters are surprised by the lack (or near-lack) of coverage for:
    • SQLite and DuckDB (seen by many as hugely important but treated as “client-side tools” rather than “serious databases”).
    • SpacetimeDB, time-series engines, vector databases, TiDB, and some immutable stores.
  • Some wish features like temporal tagging, semantic layers (e.g., Malloy), and more “Lego block” style database architectures got more systematic treatment instead of every idea becoming its own monolithic DB.

Ecosystem Trends, Monoculture, and Tooling

  • One thread laments perceived monoculture: Postgres, React, etc., becoming default answers for everything, with fewer nuanced tool choices and worsening software reliability.
  • Others ask what alternatives people actually want to see thrive; responses mention RethinkDB and more diversity in general.
  • At the same time, commenters highlight:
    • New Postgres 18 features as “fantastic”.
    • Supabase’s strong adoption among startups.
    • Ongoing innovation in vector databases for RAG/agents.

Miscellaneous Notes

  • MongoDB’s lawsuit against the FerretDB company is called “disgusting” by one commenter, though the retrospective is seen as balanced on it.
  • There’s passing discussion of multi-master clustering models (e.g., Galera-like for Postgres), MMAP vs. criticism of mmap-based DBs, and quiet but impressive execution from vendors like ClickHouse, Databricks, and Snowflake.
  • Several comments praise the CMU DB group’s teaching style and materials, noting that their “Intro to Database Systems” is an undergrad-level internals course, not an app-developer SQL course.