MatrixOne v26.4.2.7 Release Notes

MatrixOne Version: v26.4.2.7 (v4.2.7)

This release isolates server-side SQL file access, fixes locking reads with LIMIT, full-text short-pattern matching, automatic-increment and window-function behavior, and restores secondary-index access and backup history handling.

New Features

  • Server-side SQL file access boundary: A new admission check covers every SQL statement that names a server-side file path. A restricted path can only be opened when the sys account runs the statement with the moadmin role active in the session. The identity is re-evaluated per statement, table privileges never substitute for it, and no permissive fallback exists. LOAD DATA LOCAL INFILE and INLINE data are unaffected. See SQL File Access.

Behavior Changes

  • Locking reads with LIMIT: SELECT ... FOR UPDATE with an explicit, prepared, or session-limit LIMIT no longer returns an obsolete row after waiting out a concurrent update.

  • Full-text short patterns: NATURAL LANGUAGE patterns shorter than 3 runes are case-folded before prefix search, so short uppercase patterns match; json_value-indexed content and patterns of 3 or more runes are unchanged.

  • Automatic increments: LAST_INSERT_ID() and the insert ID in the OK packet never publish the hidden fake primary key used by the allocator, including for FULLTEXT index tables. A user-visible auto-increment column literally named __mo_fake_pk_col still publishes its values.

  • Window functions: Full-partition aggregates are computed once and reused across the partition, and ROWS/RANGE cumulative frames reuse aggregate states. Repeated EXECUTE of a prepared window statement re-executes cleanly instead of failing on retained state.

  • Data Branch lineage: Repeated COPY ALTER inside RC transactions no longer duplicates rows or fails the next internal drop with delete table failed: query failed.

  • Secondary indexes: Composite-key predicates are usable for secondary-index access again, and prefix-index parts no longer feed base-column values; FLOAT/DOUBLE predicates stay out of serialized regular-index access.

  • View authorization: A view-only SELECT over index JOIN rewrites keeps view authorization, and CREATE OR REPLACE VIEW keeps the view’s logical identity and grants.

  • Rollback cleanup: Statement rollback no longer duplicates persisted tombstones or fails transfer cardinality checks in RC.

  • Iceberg reads: An Iceberg table with schema but no snapshots selects 0 rows instead of ICEBERG_TABLE_NOT_FOUND.

  • SHOW TABLE STATUS: Names and current-engine statistics stay correct across CN lifecycles.

  • Asynchronous export: A cancelled asynchronous export no longer hangs finalization, and failed S3 multipart writes get bounded abort cleanup.

  • Backup history: The table-ID history required by the 24-hour GCKP history check survives a fresh physical restore, so the immediate check after a restore no longer fails.

Compatibility Notes

  • The SQL file access boundary applies to server-side LOAD DATA ... INFILE, SELECT ... INTO OUTFILE, load_file()/save_file(), file table functions, external and Hive-format readers, table dump/load, backup SQL naming server-side locations, and Datalink consumers. The boundary is effective only after every SQL-serving CN upgrades to v4.2.7. See SQL File Access.

  • Asynchronous indexed Datalink sources backed by local files become remote-only. Migrate them by repointing each source to a remote service and rebuilding the index, or materialize the local content into a TEXT column with a foreground load_text load before upgrading.

  • MO_NO_DEFAULT_CREDENTIALS and NoDefaultCredentialsForETL no longer select SQL credential behavior. Dynamic object-storage providers used by SQL statements must state credentials explicitly and cannot rely on server-local auxiliary files or server credential discovery.

  • Archives written before v4.2.7 do not carry the table-ID history objects introduced for physical restore, and that history check cannot be retrofitted to them.

Full Changelog

v4.2.6…v4.2.7