cql-hpc-bottleneck-transformer

Tiny (578k parameters) Transformer that predicts CQL performance bottlenecks from cqlc-core lexer tokens plus the shape of the data (partitions, rows per partition, row bytes, tombstones, index) and statement literals (LIMIT, IN length, batch size). Part of the cqlz family.

• Output: six bottleneck probabilities (LargeResult, ServerScan, FanOut, TombstoneScan, WriteCost, LightweightTxn) and log10(cost / baseline).
• Trained on measured latencies of public workload shapes (NoSQLBench, cassandra-stress, YCSB, TSBS, KillrVideo) on ScyllaDB and Cassandra 5.0: cql-hpc-bench.
• Files: model.safetensors, config.json (Burn ModelConfig). Use with cargo install cql-ai-cli && cql-ai check queries/ --schema schema/.

Evaluation (model trained on ScyllaDB only; Macro-F1)

Test set	with data stats	without	keyword baseline
unseen worlds	0.855	0.683	0.557
unseen workload families	0.889	0.674	0.624
other engine (Cassandra 5.0)	0.802	0.624	0.516

The released weights are trained on both engines pooled (same recipe). See the repository README for definitions, thresholds and limitations (single node in Docker on one machine; no hot-key label because no signal was observable there).