HakiCC gets large throughput gains from LLM-designed concurrency control—but verifies only observed schedules

A new University of California, Irvine paper reports that an LLM pipeline can generate workload-specific database concurrency-control protocols and then optimize their implementations against measured throughput. The headline result is substantial: across five generated protocols for the paper’s TPC-C setup, the second-stage optimizer improved throughput by 50.6% on average; across five AuctionMark protocols, the average gain was 92.2%. Those are separate workload cohorts reported on October 1, 2026—not a single combined score.

What was actually tested

HakiCC has two stages. Stage 1 uses Gemini 2.5 Pro to propose and implement a protocol, with a repair loop capped at 50 iterations. Stage 2 uses Gemini 2.5 Flash in OpenEvolve for 1,000 serially evaluated iterations, allowing localized diffs rather than full rewrites. The authors report the first five protocols produced for each workload, ten in total.

The TPC-C evaluation is narrower than the full benchmark mix. It uses New Order and Payment only, split 50/50, with 1% random user aborts. The paper varies warehouses from 64 down to one across six contention levels, and separately tests four hotspot levels with eight warehouses. AuctionMark is tested at four contention levels by changing Zipfian skew over item and user access.

Every scenario runs for 30 seconds with 64 worker threads on a Java in-memory transactional engine. The machine has two Intel Xeon Gold 6242 CPUs, 96 virtual threads and 187 GB RAM. That makes the result a controlled systems experiment, not evidence about production database engines.

Optimization found bugs, not just new algorithms

The biggest gains often came from repairing implementation defects. For the TPC-C protocol SADL, the optimizer fixed duplicate sequence-number assignment, incorrectly shared schedulers and a missing wake-up that caused head-of-line blocking. SADL then improved between 52% and 406% across the six reported contention levels.

AuctionMark shows why the average needs context. TREAD regressed 21% at low contention, then gained 480%, 106% and 91% at medium, high and extreme contention. PVL gained 435% at extreme contention, while HSL declined by 4% to 14% at every tested level. The paper’s detailed result is therefore nine improved protocols out of ten, not universal improvement.

The correctness boundary matters

All ten final protocols passed the paper’s conflict-serializability checks, including banking, shared-counter and deadlock stress workloads. But the authors explicitly describe this as schedule-dependent empirical verification: precedence-graph cycle detection establishes serializability for the observed executions, not for every possible interleaving.

That distinction is the useful result. HakiCC shows that an LLM-guided search can discover and repair high-impact concurrency bottlenecks under a closed benchmark loop. It does not yet show that generated protocols can replace formal verification or production validation.