Benchmarks

These numbers measure the full oxcode pipeline — tree-sitter extraction, cross-file resolution, reconcile into oxgraph-db, and symbol queries — on the oxgraph 0.4.0 engine versus the same pipeline on the previous published engine (0.3.2). Both columns are measured fresh against the same corpus copy, so the comparison isolates the engine.

Method

• baseline: oxcode at its pre-migration commit on published oxgraph 0.3.2.
• current: oxcode on oxgraph 0.4.0 (write-through snapshot encode, Arc-shared copy-on-write write path, checksummed self-verifying container).
• Corpus: the oxgraph workspace — 673 files · 10,428 symbols · 166,101 edges.
• Harness: scripts/bench/oxcode-perf.sh in the oxcode repo — fresh copy of the corpus, cold index, reindex with no change, reindex after a one-file edit (persistence-verified by querying the appended marker symbol), and symbol-query p50 over 5 runs. --release binaries, sequential runs on an idle 16-core machine.
• Measured 2026-06-10.

Results

oxgraph workspace — 673 files · 10,428 symbols · 166,101 edges

Reading the numbers

0.4.0 is an architecture and format release — one snapshot write path, a self-checking container (per-section CRC-32C plus a table CRC, verified at bind time), subsystem-typed errors, and structural sharing on the database write path — that holds end-to-end indexing performance while restructuring the engine underneath.

• The cold-index improvement comes from the write-through snapshot encode: freeze streams sections to their final offsets instead of owning every payload and copying the whole snapshot again (~1× peak memory, was ~2×).
• The engine-level wins show in oxgraph's own criterion suites rather than these corpus metrics: a single-element write over a committed-but-unfolded overlay of 100k entries dropped from ~18 ms to ~9–10 ms (Arc-shared copy-on-write seeding), and integrity checking moved from a whole-base CRC scan to per-section verification at bind time with no open regression.

The raw tables live in oxcode BENCHMARKS.md.