[robjtede]

Successful JSON, query, and form extraction currently pays for field-path tracking. Deserialize without tracking first, then retry over the same buffered input with serde_path_to_error only on failure. This follows tokio-rs/axum#3924 (commit 853067671a38632ceab12cc6227c046152db779c).

JSON keeps its current trailing-input behavior; this does not add an end-of-input check. Body limits, content-type checks, parser choice, and detailed error mapping remain unchanged. The deser-based extractors are unchanged.

Custom deserializers can run twice, repeat side effects, or produce a different result. A successful retry is accepted. The extractor docs state this behavior, and a test covers a stateful deserializer that succeeds on retry.

Local benchmark results

Case	Before ns/op	After ns/op	Operation-rate change	95% paired interval
query-success	327.3	180.5	81.27%	79.50% to 82.96%
query-early	420.3	621.9	-32.42%	-33.02% to -31.75%
query-late	619.8	912.4	-32.07%	-32.90% to -31.18%
json-success	890.5	742.4	19.94%	18.82% to 21.08%
json-early	1056.8	1249.0	-15.39%	-15.89% to -14.88%
json-late	1179.4	1422.2	-17.06%	-17.85% to -16.19%
form-success	1061.2	918.2	15.57%	14.23% to 17.14%
form-early	1139.5	1367.8	-16.69%	-17.30% to -16.04%
form-late	1335.6	1658.4	-19.46%	-20.43% to -18.33%

Eight alternating paired rounds, each with 1 s warmup and 1 s measurement; release builds with the same compiler and lockfile. Rust 1.100.0-beta.7 on macOS. No CodSpeed or sockets.

Payload: { n: u32, s: String, b: bool }. Early and late failures are invalid values in the first and last fields of a small payload. Query measures from_query; JSON/form include request construction, buffering, and extraction. All cases check the result; errors are formatted. These are not full HTTP throughput measurements, and absolute times are not comparable across extractor types.

The result is a success/failure tradeoff: valid inputs improve; invalid inputs cost more. Bootstrap intervals describe this local session. Compiler layout and machine noise remain possible. The tested baseline is 5d2189d and predates current upstream main; the results are not from a refreshed upstream build. Existing unrelated russe edits were identical in both benchmark builds and are excluded from this PR.

Method and reproduction steps · Raw samples · Summary

Validation

• cargo test -p actix-web-lab --lib --test extractor_error_paths --locked --quiet: 147 unit tests and 4 integration tests passed.
• just clippy: workspace, all targets, all features; passed without warnings.
• Changed Rust files formatted with nightly rustfmt and the workspace config.

Regression coverage includes diagnostic paths/messages, JSON error classification, existing body limits, trailing JSON, and successful retry of a custom deserializer.