Airline delays
Airline delays: the real-data speed ladder¶
The benchm-ml airline ladder (0.1M/1M/10M rows, mixed categorical/numeric, AUC), both the campaign knob shape and Pafka's depth-10 protocol, all rows one pod. bonsai_ts_* rows are the labeled exception to the uniform ordinal-code convention (OrderedTargetEncoder pipeline; fit_s includes the encode).
campaign knobs (depth 8), fit seconds / test AUC:
| variant | 0.1m | 1m | 10m |
|---|---|---|---|
| bonsai_depthwise | 1.2s / 0.7260 | 3.8s / 0.7447 | 22.4s / 0.7468 |
| bonsai_oblivious | 0.8s / 0.7218 | 3.3s / 0.7256 | 21.9s / 0.7264 |
| bonsai_cuda_depthwise | 0.2s / 0.7265 | 0.4s / 0.7446 | 1.9s / 0.7468 |
| bonsai_cuda_oblivious | 0.7s / 0.7218 | 0.9s / 0.7253 | 2.3s / 0.7264 |
| bonsai_ts_depthwise | 1.4s / 0.7238 | 5.9s / 0.7462 | 48.9s / 0.7498 |
| bonsai_ts_oblivious | 1.5s / 0.7247 | 5.1s / 0.7326 | 48.3s / 0.7339 |
| bonsai_ts_cuda_depthwise | 0.4s / 0.7239 | 2.0s / 0.7458 | 27.0s / 0.7499 |
| bonsai_ts_cuda_oblivious | 0.9s / 0.7247 | 2.6s / 0.7326 | 27.5s / 0.7339 |
| xgb_hist | 0.5s / 0.7263 | 2.1s / 0.7430 | 17.7s / 0.7455 |
| xgb_cuda | 0.2s / 0.7250 | 0.5s / 0.7426 | 2.6s / 0.7466 |
| lgbm_cpu | 0.9s / 0.7263 | 2.3s / 0.7431 | 10.8s / 0.7463 |
| catboost_cpu | 0.7s / 0.7167 | 3.1s / 0.7234 | 17.5s / 0.7245 |
| catboost_gpu | 0.7s / 0.7167 | 0.8s / 0.7229 | 4.7s / 0.7254 |
Pafka protocol (depth 10), fit seconds / test AUC:
| variant | 0.1m | 1m | 10m |
|---|---|---|---|
| bonsai_depthwise | - | - | - |
| bonsai_oblivious | - | - | - |
| bonsai_cuda_depthwise | - | - | - |
| bonsai_cuda_oblivious | - | - | - |
| bonsai_ts_depthwise | - | - | - |
| bonsai_ts_oblivious | - | - | - |
| bonsai_ts_cuda_depthwise | - | - | - |
| bonsai_ts_cuda_oblivious | - | - | - |
| xgb_hist | - | - | - |
| xgb_cuda | - | - | - |
| lgbm_cpu | - | - | - |
| catboost_cpu | - | - | - |
| catboost_gpu | - | - | - |
Source: airline-2026-07.jsonl. A bonsai variant has the best AUC in every cell under both protocols, and bonsai CUDA depthwise is also the fastest fit from 1M rows up under ordinal encoding; XGBoost-GPU keeps only the smallest cell. Evidence: benchmarks/airline-2026-07.md. Measured at d3ffcd0 (2026-07-31, pod-NVIDIA-L40S).