[tae898]

What

Adds an "If a graph batch is interrupted" section to docs/modules/ROOT/pages/how-to/bulk-loads.adoc (the page from #75). Python and embedded only; CHECK DATABASE is linked to the main ArcadeDB docs, not repeated.

What it says, and where the numbers come from

A graph_batch() load that is killed before close() returns (kill -9, default commit_every=1000, with and without use_wal=True, 40 kills in the probe) loses no committed chunk and leaves no edge that points at a missing vertex. But the engine links the edge records to their vertices (head pointers and incoming lists) in close(), so until then the committed edges exist (count(*) on the edge type counts them) and outE(), inE(), out() and MATCH reach none of them. CHECK DATABASE reports them as ORPHAN RECORD.

Measured on one kill (30,000 vertices, 70,000 edge records) on the native build 26.11.1.dev0, each step on the reopened database:

Step	Edge records	Reachable	CHECK DATABASE
nothing	70,000	0	70,000 ORPHAN RECORD
CHECK DATABASE	70,000	0	same
CHECK DATABASE FIX	70,000	0	same (relinks nothing)
CHECK DATABASE FIX DELETE ORPHANS	0	0	nothing left
then new_edges over the existing vertices + close()	90,000	90,000	0 warnings

Re-running the same load with the same keys: create_vertices fails with DuplicatedKeyException on a unique key and changes nothing; re-running only the edges without the delete gives 160,000 edge records for 90,000 distinct edges (70,000 old orphans stay and keep being reported). The section therefore recommends FIX DELETE ORPHANS, then an edge-only reload.

One sentence says that splitting a long edge load over several closed batches leaves only the open batch unlinked after a kill. The probe did not measure that, and the page says so.

Check

Docs only. The new section has no hard-wrapped prose and the link target was fetched before it went in.

🤖 Generated with Claude Code