ReadingResearchRadarInvestment frameworkAI Company Primary Sources
Sign in / Sign up
← Back to reading
NewsMaterial3 min readDisconfirmAI Editorial

Postgres Core Committers Number Only About 50; pg_lake Developer Craig Kerstiens Says AI-Generated Code Is Exhausting Review Capacity

In Episode 688 of The Changelog podcast, Postgres ecosystem developer Craig Kerstiens points out that Postgres's strategy of keeping its core deliberately lean and prioritizing new features in the extension layer has made it the most popular open-source database, but also means only about 50 committers are responsible for reviewing core code. The massive influx of AI-generated code contributions is exceeding the review capacity of this tiny team, while the ratio of application developers to database professionals has risen from 10:1 twenty years ago to nearly 1000:1 today.

2026-10-28 23:00 GMT+8

Postgres's success rests on a seemingly contradictory design choice: the core engine is deliberately kept "boring," with new features pushed primarily into the extension ecosystem. This strategy allowed Postgres to grow from an academic project into the most widely deployed open-source relational database globally, but it also means there are very few guardians of the core code. In Episode 688 of The Changelog podcast, Craig Kerstiens, developer of the pg_lake project, warns that this structure is facing unprecedented pressure in the AI era.

Kerstiens describes a specific numerical relationship: fewer than 100 people directly participate in Postgres core development, and only about 50 actually hold commit permissions. Meanwhile, the ratio of application developers using Postgres to database professionals has risen from 10:1 twenty years ago to 100:1 today, and is approaching 1000:1. This means the vast majority of Postgres users lack the ability to review core changes, while the team of maintainers who possess this capability is extremely small.

The influx of AI-generated code exacerbates this imbalance. Kerstiens argues that AI-generated C code is exhausting the review capacity of these roughly 50 committers. Postgres's contribution process still relies on mailing lists, without a pull request mechanism; every line of code entering the core requires human committers to review it line-by-line and understand its design decisions. When AI tools batch-generate patches that look reasonable but lack architectural understanding, the review burden does not grow linearly but expands in ways far beyond the maintainers' processing capacity.

Sign in / Sign up
ReadingResearchRadarInvestment frameworkAI Company Primary Sources

This judgment is not isolated evidence. Tom Lane, a Postgres core committer, independently described the same phenomenon in a thirty-year anniversary interview published by Snowflake: "We are receiving a huge number of AI-generated bug reports and security reports... a significant portion of which is slop because the tools do not understand basic architectural concepts." Lane also revealed that the project recently convened over twenty core members to specifically discuss whether to accept AI-generated submissions. The consensus reached so far is that a human must still be accountable for every line entering the codebase and able to explain every decision within it—but this has not yet become formal project policy.

Kerstiens's argument reveals a broader open-source governance dilemma: Postgres's "boring core" strategy was a shrewd engineering choice in the pre-AI era—it lowered the barrier to contribution, attracted many extension developers, and maintained core stability and auditability. But when AI drastically reduced the cost of generating code that "looks like a reasonable contribution," the review bottleneck of this strategy was exposed. The prosperity of the extension ecosystem does not depend on the bandwidth of core committers, but any change touching the core must still pass through the review of those ~50 people.

It should be noted that this episode has not yet released a transcript; the specific numbers and arguments above come from secondary summaries and show notes of this episode by The Podcast Summary. Kerstiens's 1000:1 ratio is his personal observation rather than strictly statistically measured data. The figure of approximately 50 committers aligns with the email-driven, self-selecting small-team governance model described in Tom Lane's interview, but the exact number fluctuates over time.

The practical implication for readers is: if you are evaluating the impact of AI-assisted programming on open-source infrastructure, Postgres provides a concrete stress-test case. It demonstrates that the true bottleneck for AI-generated code is not the ability to write syntactically correct patches, but the ability to pass through deep architectural review by a tiny team. This differs from the broad discussion of "whether AI can replace programmers"; it is a quantitative issue regarding review bandwidth and governance structures.

Sources

  • changelog.com↗
  • thepodcastsummary.comThe Podcast Summary↗
  • snowflake.com↗
Continue researching →