[tae898]

What the docs now say

The pages describe the behavior of the 26.11.1.post1 and 26.11.1.post2 builds, which the current docs do not cover.

how-to/databases-and-queries.adoc

• New section "Reading large results": to_columns(), to_arrow(), to_dataframe() and to_polars() build no dict per row, and both packages decode the same column frames, so one query gives the same column types and values on arcadedb (native) and arcadedb-jvm. The section covers the optional libraries, batch_size and columns, the memory held while frames merge, null handling for each call, the form of a vector cell for each call, dates and decimals, and the result with no rows.
• to_dataframe() keeps an integer beyond 2^53^ exact when it sits next to a floating value (the object column), also when the integer and the floats are in different frames of a large result (to_dataframe: an integer above 2**53 next to a float is rounded by pandas, the other readers keep it exact #340). The two reads that still round it are named: to_dataframe(convert_types=False) on arcadedb-jvm, and the fallback through to_list() when the Java bridge classes are missing. The limit that stays is stated too: an integer column with a null is float64 with NaN in to_columns() and to_dataframe(), so it rounds there, while to_arrow() and to_polars() keep it exact.
• The to_list() bullet points to the new section for large results. The paragraph on record wrappers states how many calls into the library save() and Vertex.new_edge make on the native build.

how-to/bulk-loads.adoc

• Defaults table: insert_columns and insert_dataframe / insert_arrow rows, and the note that a vertex type loaded through the DataFrame and Arrow calls now commits every 10,000 rows on both packages.
• New sections "Columnar loads", "Parallel and async inserts" and "Vertex types in insert_dataframe and insert_arrow". They state which columns stay columnar on each package, which ones take the row path, what a failed parallel load leaves behind (the whole open writer batch is abandoned, nothing is rolled back, count the stored rows after the error), and the retry rule for a NeedRetryException.
• The create_vertices fallback paragraph now separates the scalars that send the whole call to the older path from the values inside a list or dict that send only one chunk, and says why (the stored type of an undeclared property must not depend on which chunk held the bad value). A paragraph on the binary float32 matrix path is added.
• Interrupted-load section, "What is in the database": a killed create_vertices load on arcadedb-jvm can leave part of a call. The wording is the one fixed against the result files in the last docs commit: 240 cycles on the released 26.11.1 package, 236 with a whole number of calls and 4 with another count (the stored keys not a prefix of the input in those 4), and 2 of 240 on the 26.11.1.post1 candidate build, both a prefix. The text says that the probe has not been run on 26.11.1.post2.

how-to/vectors.adoc has a paragraph on reading many vectors that links to the new section. installation.adoc states the size of the class-data-sharing archive in the JVM wheel and that the native package has none.

Tests

tests/test_column_frames.py gets two tests that pin the new #340 statements, so the docs cannot drift from the code: the integer stays exact when the floats are in a later frame (and to_columns() returns a list of exact values), and an integer beyond 2^53^ in a column with a null is rounded in to_columns() and to_dataframe() and exact in to_arrow().

Checks

• Based on main after the 26.11.1.post2 release line was merged (Merge the 26.11.1.post2 release line into main #351), with no conflicts.
• black==26.10.1 --check, isort==9.0.2 --profile black --check-only, pyflakes, and bandit as in test-jvm.yml (--severity-level low --confidence-level low over src tests) are clean on the touched Python file.
• The docs build (npx --yes -p @antora/cli@3.2.1 -p @antora/site-generator@3.2.1 antora docs/antora-playbook.yml, the command of "Build docs site", failure_level: warn) exits 0 with no warnings, and the new anchors reading-large-results and vertex-loads render.
• The tests run in CI, not in this branch's local checks.

Refs #340

🤖 Generated with Claude Code