Oxigraph (native)

Single Rust binary, RocksDB-backed. Prebuilt Linux binary from GitHub releases.

Install + load

mkdir ~/oxigraph && cd ~/oxigraph
curl -sL -o oxigraph https://github.com/oxigraph/oxigraph/releases/download/v0.5.8/oxigraph_v0.5.8_x86_64_linux_gnu
chmod +x oxigraph
./oxigraph load --location ~/oxigraph/data --file ~/data/dblp.ttl --format nt --lenient

• Import time is "till ready", which includes RocksDB compaction. DBLP-core: 329 s parse @ 1.70 M t/s + ~243 s compaction = 572 s, 43 GB index, 561,477,456 triples. The engine isn't queryable until compaction settles.

Serve + query — two gotchas

Oxigraph (1) has no server-side query timeout (can't cancel a running query, issue #1336) and (2) has no COUNT fastpath (a full COUNT(*) over 561 M takes ~224 s). On DBLP-core it times out on ~⅔ of queries and a DISTINCT COUNT can blow memory. So:

• Run it memory-capped under systemd (MemoryMax), so a runaway query gets cgroup-killed instead of taking down the box (an uncapped run alongside other servers OOM-locked the host once).
• Run it alone on the box (the cap assumes the whole box is Oxigraph's).
• Use a per-query restart-on-failure wrapper (mirrors their util/oxigraph-helper.sh): 1 run per query, 180 s timeout (their Oxigraph value); on timeout/error, record it and restart the server before the next query. This is a documented deviation from the warmup+median-of-3 protocol — read Oxigraph's results as completed-vs-timed-out, not directly comparable.

sudo systemd-run -p MemoryMax=52G -p MemorySwapMax=0 --unit=oxbench --collect \
  ~/oxigraph/oxigraph serve-read-only -l ~/oxigraph/data -b localhost:7878
# endpoint: POST http://localhost:7878/query

See ~/run_oxigraph_safe.sh (the wrapper used for the run) — it writes oxigraph.tsv directly in the runner's TSV format. DBLP-core: 36/105 completed within 180 s, 69 timed out.