cqlz-gym-policy

Small (71,745 parameters) Transformer policy that selects Cassandra table designs, trained by REINFORCE against cqlc-core as an exact verifier. Part of the cqlz family.

• Input: per-candidate features of a design problem (cost, marginal gain, coverage, progress; 8 features).
• Output: logits over candidate tables plus a STOP action; invalid actions are masked by the environment.
• Files: model.safetensors, config.json (Burn PolicyConfig).
• Use: cargo install cqlz-gym && cqlz-gym solve workload.toml --policy hf.

Evaluation (synthetic workloads, cost model calibrated on cql-hpc-bench)

On 80–150-query workloads the policy's argmax rollout (0.16 s) is within 0.6 % of the optimum found by an exact MILP (which solves every instance in milliseconds); greedy is 11.7 % above and the branch-and-bound incumbent 7.8 % above. So the policy is a fast approximation, not better than the ILP. On a real ScyllaDB the best design raised throughput 8.2 % (sensors) and 1.4 % (shop) over the naive design. See the repository README for the full table and caveats.