March 13, 2026
Threads, tears & 32GB fears
Kovan: From Production MVCC Systems to Wait-Free Memory Reclamation
Kovan promises no more 3am memory spills; Rust crowd raves and roasts
TLDR: Kovan pitches a wait-free memory system so one stuck thread can’t bloat apps, aiming to end 3am outages. Commenters split: on-call engineers cheer fewer wake-ups, skeptics defend existing tools and call it overkill, while meme lords riff on band jokes and “lifetimes as lifeboats.”
After years wrangling high-speed data in Lever and Callysto, the creator dropped Kovan—a bold promise to end the 3am problem. Translation: instead of “lock-free” tools where the system moves forward but a single stuck thread can balloon memory (hello 32GB surprise), wait-free means every thread finishes in a bounded time, so cleanup actually happens. The post name-checks the popular crossbeam-epoch and says the quiet part out loud: stalls there can freeze reclamation for everyone, and ops folks pay the pager tax.
The comments lit up. On-call engineers cheered, swearing allegiance to anything that lets them sleep. Crossbeam loyalists called it “overengineering” and fired back with “fix your design, not your tools.” The “don’t use lifetimes as lifeboat” line became a meme, while the “Enter Shikari” header spawned band jokes and “thread mosh pit” GIFs. Claims of wins on Apple M-series Macs drew side-eye—“benchmarks aren’t production”—but TLA+ formal verification (think: math proofs your system won’t explode) earned cautious respect. The community’s split: half craving fewer pages and bounded behavior, half warning we’re swapping practical pain for academic purity. Fans pointed to an ecosystem in progress and research roots like “Crystalline,” while skeptics asked for multi-tenant, Linux server numbers before declaring victory.
Key Points
- •Production experience with Lever (MVCC/STM) and Callysto revealed limitations in lock-free, epoch-based memory reclamation at scale.
- •Epoch-based reclamation (e.g., crossbeam-epoch) is lock-free but not wait-free; a single stalled thread can block reclamation across the process, causing unbounded memory growth.
- •The article argues for wait-free memory reclamation to guarantee bounded operation steps, eliminating starvation and reliance on scheduler fairness.
- •Research such as “Crystalline: Fast and Memory Efficient Wait-Free Reclamation” is cited as inspiration for the wait-free approach.
- •The post introduces Kovan and outlines planned sections on its design, performance vs crossbeam-epoch on Apple M-series Macs, an ecosystem, and formal verification with TLA+.