Expand Rust AEL fast path with native control lowering

aerospike_sdk/ael/parser.py

@@ -56,6 +56,7 @@
     MapValuePart,
 )
 from aerospike_sdk.ael.filter_gen import FilterGenerator, IndexContext, ParseResult
+from aerospike_sdk.ael._rust_fastpath import try_parse_ael_rust
 
 
 class _AELParseErrorListener(ErrorListener):
@@ -163,6 +164,9 @@ def parse_ael(expression: str, *args: Any) -> FilterExpression:
     if _parser is None:
         _parser = AELParser()
     placeholder_values = PlaceholderValues(*args) if args else None
+    rust_result = try_parse_ael_rust(expression, placeholder_values)
+    if rust_result is not None:
+        return rust_result
     return _parser.parse(expression, placeholder_values)
 
 

docs/benchmarks/ael-rust-fastpath-results.md

@@ -0,0 +1,68 @@
+# AEL Rust Fast Path Results
+
+This benchmark note captures the current behavior and performance of the Rust-backed AEL fast path after native lowering for `let(...)`, `when(...)`, and variable references.
+
+## Verification
+
+- `cargo test --manifest-path rust/ael_rust_ext/Cargo.toml --target-dir /tmp/ael_rust_ext-target`: `5 passed`
+- `pytest tests/unit -q`: `1581 passed`
+- `pytest tests/integration -q`: `979 passed, 37 skipped, 2 xfailed`
+
+## Parser Benchmark
+
+![AEL parser throughput](ael-rust-parser-throughput.svg)
+
+![AEL parser latency](ael-rust-parser-latency.svg)
+
+### Repo expression corpus
+
+- Raw corpus size: `868`
+- Supported by Rust fast path: `750`
+- Raw skips: `118`
+- Skip classification: `28` placeholder-only cases with no values supplied, `90` expressions that Python itself rejects
+- Real valid-expression fallbacks: `0`
+- Mismatches vs Python: `0`
+- Python parser throughput: `4,798.66 expr/s`
+- Rust fast path throughput: `12,988.76 expr/s`
+- Python parser time: `208.39 us/expr`
+- Rust fast path time: `76.99 us/expr`
+- Speedup: `2.71x`
+
+### Literal `parse_ael(...)` callsites
+
+- Corpus size: `596`
+- Supported by Rust fast path: `596/596`
+- Mismatches vs Python: `0`
+- Python parser throughput: `4,496.22 expr/s`
+- Rust fast path throughput: `9,842.60 expr/s`
+- Python parser time: `222.41 us/expr`
+- Rust fast path time: `101.60 us/expr`
+- Speedup: `2.19x`
+
+## Live Workload Benchmark
+
+Environment:
+
+- Aerospike server: Docker `aerospike/aerospike-server:8.1.0.2`
+- Address: `127.0.0.1:3000`
+- Workload: `RU,50`
+- Command: `python -m benchmarks.benchmark -k 100 -z 4 -w RU,50 -d 3 --warmup 0 --cooldown 0`
+
+Results:
+
+- Total TPS: `8389`
+- Read TPS: `4161`
+- Write TPS: `4228`
+- Latency: `p50=0.4ms p90=0.9ms p99=1.4ms p99.9=1.6ms max=1.8ms`
+- Peak RSS: `48.8 MB`
+
+Comparison run:
+
+- PAC async: `9016 TPS`, `p99=1.4ms`, `RSS=42.3MB`
+- PSDK async: `8486 TPS`, `p99=1.4ms`, `RSS=49.0MB`
+- PSDK sim-sync: `3007 TPS`, `p99=0.7ms`, `RSS=49.2MB`
+
+Interpretation:
+
+- The parser work is materially faster.
+- The live `RU,50` benchmark does not show a major end-to-end TPS change because it is primarily request/response bound rather than parser bound.

rust/ael_rust_ext/Cargo.toml

@@ -0,0 +1,11 @@
+[package]
+name = "ael_rust_ext"
+version = "0.1.0"
+edition = "2021"
+
+[lib]
+name = "ael_rust_ext"
+crate-type = ["cdylib"]
+
+[dependencies]
+pyo3 = { version = "0.21.2", features = ["extension-module"] }