solana-rpc-reviewlisted

# Solana RPC Review Review skill for the specific class of code [Company] ships: RPC infrastructure serving 100k+ msg/sec to trading firms, indexers, and oracles. Different review priorities than typical web code — latency at p99 matters more than features, and a 2ms regression in the hot path is a blocker. ## Hot path discipline Any code in the request/message handling path is "hot". For these paths: ### Allocations - **Zero allocations per message** is the goal in steady state. Reuse buffers, use object pools, prefer `&[u8]` over `Vec<u8>` when not consumed. - Look for: `format!`, `to_string()`, `clone()`, `Vec::new()` inside loops, `Box::new()`, lazy `String` concatenation. All suspect. - Profile with `dhat`, `heaptrack`, or `cargo flamegraph` before claiming perf is fine. - Strings: prefer `&str` over `String`, `Cow<'_, str>` when sometimes-owned. - Numbers in messages: use `Bytes` from `bytes` crate for zero-copy slicing. ### Locks and contention - `Mutex`/`RwLock` in the hot path = stop and justify. Use `parking_lot` variants (faster) or `arc-swap`/`atomic` types for read-heavy state. - `tokio::sync::Mutex` is async-aware but slower than `parking_lot::Mutex` for short critical sections. Use the right one. - Sharded locks (`DashMap`) often beat `Arc<RwLock<HashMap>>` when contention is real. But profile, don't guess. - Check for accidental serialization: futures awaiting a global resource serialize concurrent requests. ### Async correctness - **Nev