antirez built a new Redis Array data type over four months, using AI (Opus, GPT 5.x, Codex) for spec writing, code generation, and testing at production quality.
Key Takeaways
The Array type uses a three-tier structure: sparse representation, directory+slices, and a super-directory of sliced dense directories (4096 elements/slice) to handle large non-contiguous indices without huge allocations.
ARSCAN and ARPOP scan in time proportional to existing elements, not range span, which is the key performance property distinguishing this from ZSET.
ARGREP adds regex search using TRE (with antirez-patched OR-pattern optimization and security fixes) after antirez started storing markdown files in Redis arrays as a knowledge base.
AI provided leverage on two specific tasks: exhausting boilerplate like 32-bit support and fuzz-style coverage of complicated algorithm edge cases. Core design decisions remained human-driven.
The full implementation is ~5000 lines (2000 sparse array, 2000 command layer, ~500 AOF/RDB); PR is open at github.com/redis/redis/pull/15162.
Hacker News Comment Review
Commenters debated whether Array overlaps too much with ZSET; the distinction is that numerical index is semantic in Array, and dense/sparse auto-promotion is built-in rather than bolted onto ZSET internals.
The AI workflow antirez describes – spec-first, line-by-line review, AI as safety net – matches what several senior practitioners reported independently, pushing back on “vibe coding” framing.
Skepticism exists that the AI praise is partly Redis marketing for the vector/AI-use-case market, and that the 22k-line PR size makes community review difficult compared to incremental mailing-list development (Postgres model).
Notable Comments
@tibbar: Flags that a 22,000-line PR with complex feature set is hard to review; contrasts with Postgres incremental mailing-list approach.
@antirez: Clarifies actual code is ~5000 lines; the rest is tests, JSON descriptors, and the TRE dependency.
