corebreaker / arbordb / 21

88%

main: 98%

Build Type

Pull #2

circleci

Committed by corebreaker

Commit Message

Extend the write-back cache to permissions and to file edits, and add a direct typed-to-blob encoder

Rename the per-transaction DirBuffer to WriteBuffer and compile it into every build, including permissions, where a flushed directory is now sealed once at commit. A protected fresh bulk under one parent was O(N²) and re-signed the parent directory on every child link (2N Ed25519 signatures); it is now O(N) with the parent signed once — protected bulk drops from 119 to 24 µs/op at N=2000 (~4.9×) and is linear again. A freshly created directory's blob is buffered but its ACL is stamped eagerly, so a non-master writer's mid-transaction traversal check still sees it; resolve_enforced, Grab::entry_blob, Grab::verify, and the Context read helpers consult the buffer first, since a buffered a-node is the transaction's own trusted state (sealed only at the commit-time flush). External tampering is still caught, since a fresh transaction starts with an empty buffer.

Add a direct typed-to-blob encoder: AData::encode_node emits a value straight into the blob in one children-first pass (driving a doc-hidden NodeEncoder), so store<T> no longer decomposes the value into a BTreeMap-backed Value tree and then walks it a second time. The default keeps the old Value-based path (so every implementor is correct); the scalars, containers, and — in the derive — simple structs override it, with a flatten or store_with field falling back to the default. The output is byte-identical to the Value path (cross-checked). Non-permissions fresh bulk drops from 2.77 to 1.99 µs/op (−28%); the remaining gap to raw redb is the self-describing blob (field names plus offset tables), the deliberate trade that wins on reads.

Buffer edits of an existing file's value in the same write-back cache, so a burst of edits to one value in a transaction coalesces into a single re-encode and, under permissions, a single signature at the commit-time flush — the incremental cost falls to ~2.1 µs/edit... (continued)

Pull Request Pull Request #2: Optimization

Coverage Stats

898 of 950 new or added lines in 33 files covered. (94.53%)

20 existing lines in 8 files now uncovered.

8445 of 9666 relevant lines covered (87.37%)

374911.34 hits per line