“We should be more ambitious with the software we build”

Testing pgrust has revealed genuine bugs within Postgres, such as a numerical issue causing valid SP-GiST operations to fail, a format cache bug leading to incorrect JSONPath results, a pgcrypto bug that can impede subsequent operations, multiple memory safety issues in the dictionary loader, and a money overflow bug.

Michael Malis of pgrust has been sharing his work since early this year, suggesting that "We should be more ambitious with the software we build." Despite the current flood of low-quality AI-generated projects, which naturally breeds skepticism toward initiatives like pgrust, the thorough testing demonstrates significant effort.

The primary selling point for pgrust centers on performance gains. Should the project conclude, some of the demonstrated optimizations could potentially migrate into Postgres or Clickhouse, though fixing direct Postgres bugs offers a more immediate benefit.

While many dismiss such rewrite projects as "vibeslop garbage," the author notes that for builders, these ambitious attempts matter. With modern tooling, projects like pgrust, previously too large to feasibly tackle, can achieve a usable quality level through diligent testing and benchmarking.

The author suggests that Malis’s sentiment implies that people are simply not considering these advanced endeavors because they are now achievable. For instance, bottlenecks like Ripgrep or Rust compilation speed are now areas where substantial improvements are possible.

Advancements in capability are evident; for example, the author built a regex engine fork faster for specific Codex Ripgrep workloads and a Rust tool that accelerated builds, all in a low single-digit number of hours.

Furthermore, switching development environments from Mac to Linux significantly improved syscall and filesystem overhead, revealing that modern disks negate the need for traditional *nix filesystem APIs, making direct I/O viable for developer toolchains.

In the realm of AI, the author observed that while early attempts to create superhuman board game AIs were time-consuming, newer models like Fable 5.1 and Opus 5.5 show a much better grasp of these domains, dramatically cutting down development time.

Even with advanced models, the claim that "software is solved" is inaccurate. For instance, GPT-6 Astra successfully passed all registered zstd tests over 99% of the time, while also handling complex implementations like the IMAP RFC and Pandoc, tasks that earlier models struggled with.