[tae898]

Closes #229

On the arcadedb-jvm build, insert_many now pauses Python's cyclic garbage collector for the duration of a call when rows is a list or a tuple, and restores the previous state in finally (re-enabled only if it was enabled on entry). A generator, an iterator, a DataFrame or any other lazy input is not paused. The native build is not changed.

The change is the arm C of the study python-gc-ingest-ab-laptop-20261025 (HumemAI benchmark notes), adopted by the decision in DECISIONS #238 addendum on TA-2 and #239: the study's rule was written before the first timed run and C passed it. Wheel measured: arcadedb-jvm 26.11.1.dev0 from CI run 41646452646 on main 413f54ffe1; 500,000 dict rows in slices of 100,000, ABBA order, ab.py --noise, 95% CI of the median ratio against the default:

Figure	Candidate	Clean pairs	Median ratio vs default	95% CI
insert_many wall	library pause (this change)	12	0.860	[0.819, 0.903]
insert_many wall	caller gc.disable()	13	0.854	[0.797, 0.875]
whole load wall (rows built too)	library pause (this change)	12	0.876	[0.861, 0.923]
whole load wall (rows built too)	caller gc.disable()	13	0.881	[0.847, 0.895]
peak RSS	library pause (this change)	12	0.976	[0.924, 1.057]
peak RSS	caller gc.disable()	13	0.940	[0.895, 1.076]
insert_many wall	A/A control	13	1.080	[0.957, 1.105]
whole load wall	A/A control	13	1.044	[0.974, 1.084]

The shipped form follows the risk section of the study README: the guard to a list or a tuple is what keeps a generator that builds reference cycles from piling up garbage for the whole load. gc.disable() is process wide, so with overlapping calls from several threads the pause does not last for the whole of each call, and it is never left stuck off.

Code: Database.insert_many in src/arcadedb/core.py records gc.isenabled(), calls gc.disable() for a list or tuple, runs the unchanged body (moved to _insert_many_nogc) and restores in finally. Docs: the docstring and a new section on the bulk-loads page, with no speed figure. Tests: tests/test_insert_many_gc_pause.py (JVM only, listed in NATIVE_NOT_APPLICABLE) observes gc.isenabled() from inside the call for a list and a tuple, several chunks, the parallel mode, an exception, a collector the caller had disabled, and a generator, an iterator and a generator expression.

CI (test-native.yml dispatched, linux-x86_64 and the other legs): tests only, commit 1d64df8 on the branch perf/native-point-op-glue-tests-first: https://github.com/ArcadeData/arcadedb-python/actions/runs/41715698870 ; with the cuts, head 2c1aa6a: https://github.com/ArcadeData/arcadedb-python/actions/runs/41715725208 . Results are added here when the runs finish (the macOS image build leg gates the test jobs and was still running at the time of writing).

Edited 2026-11-05: the closing reference was #235 (an issue another PR closed); this PR closes #229.

🤖 Generated with Claude Code

[tae898]

Test runs for this PR, both read on 2026-11-04:

Run	Branch	Code	Result
41715725208	perf/native-point-op-glue (head 2c1aa6af83)	this PR	every job green
41715698870	perf/native-point-op-glue-tests-first (head 1d64df8359)	main plus only the new test file	test_native_point_op_glue.py: 41 failed, 41 passed; every other file passes (linux-x86_64 py3.12 parity: 2039 passed, 41 failed, 0 errors)

The 41 tests that fail on main are the ones that check the changed path; the 41 that pass on main check behaviour that must not change. The tests-only branch is deleted after this read, and its remaining macOS legs were cancelled to free the shared runners.

[tae898]

Timed A/B (native wheel; arm A = main's CI native wheel, arm B = the same tree with only this PR's files, A/A control; laptop timed lane, tools/ab.py --noise, ABBA, round 0 dropped). Per-operation time, B/A median with 95% CI (below 1 = this PR is faster):

operation	role	B/A	A/A
lookup_by_key	gating	0.777 [0.772, 0.784]	1.001 [0.995, 1.007]
out edges of a vertex	gating	0.843 [0.836, 0.849]	1.001 [0.996, 1.007]
in edges of a vertex	gating	0.842 [0.837, 0.845]	1.001 [0.994, 1.014]
ends of an edge	gating	0.893 [0.887, 0.897]	0.999 [0.995, 1.005]
point SELECT (statement path)	control	0.998 [0.996, 1.002]	0.999 [0.992, 1.006]
both-direction edges	info	0.848 [0.843, 0.849]	1.005 [0.993, 1.009]
lookup_by_rid	info	0.991 [0.988, 1.000]	1.006 [0.994, 1.007]

Clean pairs per figure: 12; every run passed its arm check, and the A/B output differential passed before the run. By the rule written before the run this is a SHIP: every gating figure is below 1.00 with the control flat.

The merge waits for the end of the current batch of timed studies that pin main (merging now would move their pins); the PR leaves draft at that point.