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

benchmarks: BENCH_INSTRUMENT=2026-11 turns on the default-off switches and the agreed headline changes #412

Closed

Description

@tae898

The next instrument (BENCH_INSTRUMENT=2026-11, added by the instrument selector) still measures exactly what 2026-10 measures, because the default-off switches are all off and the agreed headline changes are not built. This issue makes 2026-11 the behaviour; unset BENCH_INSTRUMENT must stay byte-identical to 2026-10.

What turns on under 2026-11: the sparse lane warm-up for every engine (100 untimed searches) and the idle-state hold for timed cells (target 0 microseconds); the dense id index on the seven engines that lack one; the official arcadedb-driver (sync) as the served ArcadeDB headline client, with the lean HTTP client as a labelled twin cell; l4 q_high and q_groupby read on each engine's documented best path for large results (ArcadeDB to_arrow, DuckDB to_arrow_table, SQLite fetchall) with to_list / fetchall timed beside it; and embedded ArcadeDB arms no longer pin -Xms to -Xmx.

Rules: an explicit environment variable keeps overriding the default in both instruments, rows stamp what they used, BENCH_ARCADEDB_TS_NDJSON stays off (measured mixed), and each switch gets one test under 2026-10 and one under 2026-11. CAMPAIGN, PROTOCOL, FAIRNESS and PAGE-SPEC describe the new defaults.

Not built here: the PostgreSQL wire as a sensitivity row (the harness has no adapter for it), and the official driver for bulk loads (they stay on HTTP batch).

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions