Skip to content
This repository was archived by the owner on Nov 21, 2026. It is now read-only.

Benchmark page: bindings cutoff 26.10.2, index-before-load sentence, three declarations - #397

Merged
tae898 merged 2 commits into
mainfrom
export-page-notes
Oct 21, 2026
Merged

tae898 merged 2 commits into
mainfrom
export-page-notes

Conversation

@tae898

@tae898 tae898 commented Oct 21, 2026

Copy link
Copy Markdown

Closes #396

Report-time changes to the benchmark page generator (benchmarks/experiments/export_web.py). No measurement code, stage file or stored row changes.

_BINDINGS_FASTER_FROM goes from 26.11.1 to 26.10.2. #277, #281, #286, #295 and #320 are all ancestors of the 26.10.2 tag, the release run for that tag built the PyPI wheel, and the wheel's Python sources are byte-identical to the tag's, so a 26.10.2 embedded row already contains the bindings speedup and must not be told it predates it. The 26.10.1 rows still get the line, and the ts_doc out-of-memory declaration, keyed on the same constant, still describes only wheels below 26.10.2, which is where #295 landed.

An index defined on an empty type before the load times as a few hundredths of a second and read as a free index (arcadedb_ts_doc shows an index s of 0.04 s against a 27 s load). The page never read index_before_load. _index_before_load_notes now prints one sentence under each table with an index s column naming the arms that define their index before the first row: the arms whose rows carry the flag (SurrealDB), plus _INDEX_BEFORE_LOAD_REPORTED for the arms whose adapters do not set it (ArcadeDB's embedded and served document time-series arms, ArangoDB's time-series arm, SurrealDB's embedded cross-model arm). The rule is the same for every engine. Arms that also build a larger index after the load (ArcadeDB's cross-model arm: Product(pid) first, the vector index after) print a real index s and are not listed. I checked every index_timer site in the lane adapters; the graph and sparse tables print no index s column, so they show nothing to misread.

Three declarations. OCT_PROSE["GLOBAL"]["index_order"] says ArcadeDB's maintainers advise creating indexes after a bulk load and the ArcadeDB rows create most of theirs before it. jvm_heap says the JVM engines run with the initial heap pinned to the maximum heap, against ArcadeDB's guidance for measuring memory. _gav_supertype_note declares that the Graph Analytical View lacks the Message supertype while the rows are 26.10.x, because listing it returns a wrong count for one graph query there, and it refuses a 26.11.1-or-later row whose gav_types still lacks it, so the 26.11.1 measurement cannot be published with the stale list. That guard sits in the export, not in l2_graph.py, because a harness change needs its own pull request; it carries a REMOVE WHEN line.

Not done here: the tracked results/web_benchmarks*.json payloads are not regenerated, since they are built from the campaign's run log, which is not in the repository, and the next stage landing regenerates them.

Checks run, from benchmarks/experiments with uv run: the whole experiments test directory (765 passed, 14 skipped), including the new test_index_before_load_note.py and test_gav_supertype_note.py and a new case in test_next_remeasure_line.py; the repository's pre-commit on the six changed files.

🤖 Generated with Claude Code

tae898 and others added 2 commits October 21, 2026 19:11
…three declarations

The "newer Python bindings" line is keyed on 26.10.2, the first wheel that carries the bindings changes (all five PRs are in the 26.10.2 tag and the PyPI wheel's sources are byte-identical to it). Arms that define their index before the first row are named in one sentence under each table with an index s column, from the rows' index_before_load flag plus a report-time list for the arms whose adapters do not set it. Two global declarations (index order, pinned JVM heap) and one for the graph view's Message supertype, which the export refuses on 26.11.1 rows that still lack it.

Closes #396

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
….11.1 re-measure and go on in the next instrument version, with every touched engine re-measured (#396)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@tae898
tae898 merged commit 27a80d3 into main Oct 21, 2026
3 checks passed
@tae898
tae898 deleted the export-page-notes branch October 21, 2026 10:33
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Benchmark page: bindings cutoff, free-looking index timing, and three declarations

1 participant