[tae898]

This is the port of humemai/arcadedb-embedded-python#528 (package arcadedb, no "embedded" names).

insert_many sends each 10,000-row chunk to the engine as one JSON text. Until now json.dumps raised on a datetime, date, Decimal, bytes, a numpy scalar or a pandas NA/NaT, and one such value anywhere in a chunk sent the whole chunk down the per-row path (new_document + set per property + save, about 10 JPype crossings per row). Real tables nearly always carry a timestamp or a decimal, so most loads never reached the bulk path (backlog rows N-04, J-04, J-05).

What changes

• A value that sits directly in a row now crosses inside the JSON as a tagged value {"\u0001": [kind, payload]}: T (datetime, as the UTC wall clock the per-row path stores), D (date), n (Decimal), b (bytes, base64). DocumentBatcher.fill decodes the tag to the object convert_python_to_java builds for the per-row path (LocalDateTime, LocalDate, BigDecimal, byte[]). Numpy scalars cross as plain numbers (np.bool_ as a bool), pandas.NA/NaT/np.datetime64('NaT') as null, pandas.Timestamp and np.datetime64 as a naive datetime (microseconds, as the per-row path keeps them).
• A value the bulk path still cannot carry sends its own row down the per-row path, not the 10,000 rows around it: NaN and Infinity, an integer beyond 64 bits, a datetime/Decimal inside a list or dict, a non-str dict key, a lone surrogate, a datetime.time, a numpy longdouble or timedelta64, a non-finite Decimal. Runs of carriable rows still go in bulk, and every row is inserted in the order given. In parallel=True the uncarriable rows are written synchronously after the writers drain (the writers never kept order).
• json_bulk_dumps (used by GraphBatch) is unchanged; only insert_many uses the new json_bulk_dumps_tagged, because only DocumentBatcher.fill decodes the tags. A user dict keyed by the tag marker is refused by the bulk path, so it stays a map.
• Not in this PR: insert_dataframe/insert_arrow/insert_columns with a null in a numeric column (the other half of J-05) still builds an Object[] element by element. That is a separate Java change (a validity mask), so it gets its own PR.

Stored values

tests/test_insert_many_mixed_values.py writes the same rows through the per-row path and through insert_many, into a declared type (DATETIME, DATE, DECIMAL, BINARY, LONG, INTEGER, DOUBLE, FLOAT, BOOLEAN properties) and a schemaless one, and compares the stored value and its Python type for 17 value kinds (naive and UTC+9 datetime, date, Decimal with and without exponent, bytes, bytearray, np.int64/int8/uint32/float64/float32/bool_, None). It counts Database.new_document calls to prove which path ran, and also covers pandas.Timestamp (naive with nanoseconds, aware), pandas.NA/NaT and np.datetime64 (the per-row path raises for these; the bulk path stores null / a naive datetime), a 10,000-row chunk with one datetime, an uncarriable row in the middle of 1,000 (exactly one new_document call, stored rows equal the per-row path's), insertion order around per-row runs, commit_every across chunks, rollback on a failing row, parallel=True, and the tag lookalike. The suite failed 15 of 20 tests before the change.

Untimed 1M-row figure

One run each (not through ab.py, so no claim), fresh process, 1,000,000 rows of six properties with a Python datetime per row, row building inside the timed call, JVM -Xmx4g, through quiet_build: insert_many took 31.2 s before (1,000,000 new_document calls, the whole load on the per-row path) and 7.4 s after (0 new_document calls). The same rows without the ts column take 4.1 s, so the datetime itself now costs about 3.3 s per million rows (JSON text plus LocalDateTime.parse) instead of 27 s. Measured on the humemai fork (humemai/arcadedb-embedded-python#528) with the same change.

Local checks

• tests/test_insert_many_mixed_values.py 20 passed, plus test_bulk_insert.py, test_bulk_json_paths.py, test_float_vector_bulk.py, test_numpy_support.py, test_conversion_fixes.py, test_graph_batch.py, test_insert_columns.py, test_async_executor.py, test_core.py, test_bridge_fixes.py, test_type_conversion.py, test_docs_examples.py, test_check_test_skips.py (all pass) against the engine jars and JRE of the humemai main CI wheel with this branch's Python files and the bridge classes compiled into a copy of the bridge jar (171 passed for the five suites listed first)
• black, isort and the repository pre-commit hooks pass, bandit reports nothing at medium or above
• The full suite and the wheel build are left to CI

Docs changed in this PR: the insert_many docstring (what is carried, what still goes per row). This repository keeps its API reference in Antora modules that do not describe the insert_many fallback, so there is no other page to update.

🤖 Generated with Claude Code

[tae898]

Moved to #82: the personal fork was retired, so the branch now lives on ArcadeData/arcadedb-python as mixed-chunk-bulk (same commits). Continue there.