starknet_committer,starknet_patricia,starknet_patricia_storage: hoist per-node db key prefix in index-layout serialization

[gkaempfer]

Summary

Follow-up performance review of #14850 (which removed a redundant .to_vec() copy in create_db_key), scoped to the same hot path: Patricia trie DB key construction, which runs once per node per block during block commitment.

In the index-layout storage backend, IndexLayoutStarknetStorageValue::get_static_prefix (crates/starknet_committer/src/db/index_db/leaves.rs) rebuilds a fresh 32-byte Vec<u8> from a contract's address for every node in that contract's storage trie (inner nodes and leaves alike) — even though every node serialized in one FilledTree::serialize call, or deleted in one serialize_deleted_nodes iteration, shares the same contract address. For a contract with thousands of modified storage slots in a block, this recomputes and heap-allocates an identical prefix thousands of times instead of once.

Facts layout doesn't have this problem — its prefix depends on node type (inner vs. leaf), not just on the key context — so it's left untouched.

Change

• starknet_patricia_storage::db_object: added HasDynamicPrefix::static_prefix_hint(key_context) -> Option<DbKeyPrefix>, a hook that returns the prefix without an instance when it depends only on key_context (Some, via the blanket HasStaticPrefix impl — covers index layout) or None when it's node-data-dependent (facts layout's manual HasDynamicPrefix impl doesn't get the blanket override, so it stays None, and its get_db_object ignores the hint — unchanged behavior).
• starknet_patricia::db_layout + filled_tree::tree: NodeLayout::get_db_object gained a prefix_hint: Option<&DbKeyPrefix> parameter; FilledTreeImpl::serialize computes the hint once before its per-node loop and reuses it for every node.
• starknet_committer::db::index_db::db: IndexNodeLayout::get_db_object uses the hint when present; serialize_deleted_nodes (a separate per-contract loop) now computes the prefix once per contract instead of once per node index.
• create_db_key now borrows its prefix (&DbKeyPrefix) instead of taking it by value, since callers with a precomputed prefix no longer need to move it per node; all call sites updated.
• Added unit tests (starknet_patricia_storage::db_object_test) directly pinning the invariant the optimization relies on: the hoisted hint equals the per-instance prefix, and two objects sharing a key context but differing in data produce identical db keys via both paths.

Test plan

• cargo build -p starknet_patricia_storage -p starknet_patricia -p starknet_committer -p starknet_transaction_prover
• SEED=0 cargo test -p starknet_patricia_storage -p starknet_patricia -p starknet_committer -p starknet_transaction_prover — all green, with two known-unrelated exceptions confirmed pre-existing on a clean origin/main checkout:
  • starknet_transaction_prover's test_compiled_class_v1_to_casm_round_trip fails in this sandbox because it needs to download the Cairo 1 compiler over a network the sandbox doesn't have; reproduces identically on origin/main.
  • starknet_patricia_storage's test_storage_concurrent_access::case_2_mdbx_storage uses a hardcoded /tmp/test_mdbx_storage path with no cleanup between runs; it only fails when the suite is re-run repeatedly against leftover on-disk state from a prior run in the same environment. Passes cleanly otherwise.
• unset CI && scripts/rust_fmt.sh — no diff
• Reviewed by an independent Opus pass, which added the db_object_test.rs unit tests above and confirmed no blocking issues; a second independent Opus pass reached the same verdict.

🤖 Generated with Claude Code

---

Generated by Claude Code

[cursor]

PR Summary

Medium Risk
Changes how storage keys are assembled on the block-commitment serialization path; keys must stay byte-identical, which the new tests are meant to guarantee.

Overview
Performance follow-up on Patricia trie DB key construction during block commitment: index-layout storage was rebuilding the same contract-address-derived prefix for every node in a contract storage trie (and again when serializing deleted nodes).

Adds HasDynamicPrefix::static_prefix_hint (and DBObject::db_key_from_prefix) so callers can compute a prefix once when it depends only on key_context. FilledTree::serialize hoists that hint before its per-node loop; NodeLayout::get_db_object takes an optional prefix_hint. Index layout uses the hint; facts layout still derives prefix per node (inner vs leaf) and ignores the hint.

create_db_key now takes &DbKeyPrefix; get_node_index_db_key_with_prefix supports deleted-node key batching with one prefix per contract. New db_object_test asserts hoisted keys match the old per-instance path.

^{Reviewed by Cursor Bugbot for commit 911cc8f. Bugbot is set up for automated code reviews on this repo. Configure here.}

[reviewable-StarkWare]

This change is