2026

Category

Feature

Description

Related documentation

New features and optimizations

Vector engine

• Optimized the exact recall computation performance of Nova autotune on large-scale datasets.

• Fixed the issue where NOVA partial indexes did not filter data by index predicates during the stock build phase.

• Fixed the recall loss of NovaM SQ8 caused by uneven quantization precision after PCA was enabled on L2 datasets.

• Optimized dataset identification and noise parameter selection in the no-test-set mode of Nova autotune.

• Fixed the issue of missing batch ground truth functions for Nova autotune after upgrading existing FastANN 1.2 instances.

• Fixed the issue where ivfflat.probes in the pgvector-compatible API could only be set and displayed but did not affect the actual search nprobe of NOVAD.

• Fixed the issue where NOVA vector search did not execute filter conditions in some statistics scenarios.

• Fixed the issue where insufficient NOVAD delta capacity could cause missing index entries and persistent write waits.

• Fixed security vulnerabilities related to pgml model deserialization, file access, and external dataset loading.

None

Stability

Fixed the issue where pg_stat_all_tables did not correctly count HOT updates.

None

Optimizer

• Added single-table Rows hints and extended Rows hint support in the ORCA optimizer; added fuzzy (wildcard) Leading and spanning join method hint syntax; added hint rule tables and management functions based on table sets. Optimized Rows hint semantics for inheritance and partitioned tables (parent table total pushdown to rebuild member paths and incremental propagation of member real names).

• Added EagerAgg control enhancement GUCs to support forced two-phase splitting, join key distribution optimization, recursive pushdown depth control, and NDV row count estimation fixes.

• Optimized the CTE inlining logic of the ORCA optimizer. Through cost-driven selective inlining decisions, CTEs that should be inlined and CTEs that should be shared are precisely distinguished in TPC-DS-like analytics scenarios.

None

Execution engine

• The compute engine supports intervention for Skew value Join, providing a general solution for corner performance problems caused by skew on a specific value.

• Added Beam support for Segment Key In Light & storage elasticity.

None

Bug fixes

Stability

• Fixed the "could not find CTE" error caused by double adjustment of ctelevelsup for chained CTEs in views in CTE pullup scenarios; also fixed the deserialization failure caused by mismatched assertOneRow/encodetype element names in readfuncs.c.

• Fixed the global variable concurrency safety issue in LATERAL subquery pullup rewriting, the incorrect quals processing on the non-nullable side of outer joins, and the segmentation fault caused by NULL context.

• Fixed the SIGSEGV crash of generate_union_paths when processing SET operations.

None

Optimizer

Optimized the ORCA partition pruning construction logic to avoid incorrect join order caused by partition pruning logic.

None

Scaling

Fixed the issue where the gpexpand packaging template did not preserve sparse file attributes, causing abnormal disk space usage.

None

Execution engine

Fixed the coredump issue when Beam subscription tables performed index scans.

None

Real-time materialized view

Fixed the issue where IMV creation did not block pure time functions in subqueries; fixed the crash caused by the check same logic when CREATE OR REPLACE carries new reloptions.

None

Category

Feature

Description

References

New features and optimizations

Real-time materialized views

Fixed an issue where reloption inheritance failed during CREATE OR REPLACE for nested incremental materialized views. A new Hybrid index decision mode is added, which selects a better apply-delete index by comparing the actual combined selectivity of user-specified columns and auto-candidate columns.

None

Vector engine

• Optimized Motion network transmission overhead for vector distance queries in join scenarios by precomputing distance values before redistribution.

• Unified the Nova Delta Flush monitoring interface to use nova_delta_flush_history() for querying database-level refresh history.

• Fixed an issue where the explicit noise_scale parameter was overridden by dimension defaults in Nova autotune for 768/1024-dimensional vector scenarios.

• Nova now supports CREATE INDEX CONCURRENTLY.

• Optimized Nova Delta Flush and Backpressure default parameters to shorten the incremental vector refresh cycle and reduce write backpressure latency.

• Optimized the NOVAD autotune parameter search strategy to improve configuration quality for high-recall targets and reduce tuning time.

• Optimized the Nova autotune API, parameter grid, and QPS selection logic to improve consistency of automatic tuning results.

• Nova index auto-tuning is supported.

• Fixed an issue where the novam_reorder_graph graph reordering UDF reported errors on distribution tables because the master node had no data. The UDF now gracefully skips when no data is available instead of throwing exceptions.

• Nova now supports ARM platforms.

• Result and Subquery now support dictionary generation and pushdown.

None

High availability

Fixed a bug where slow query memory usage showed negative values.

None

Bug fixes

Execution engine

Fixed an issue where a missing expression compile callback in the JIT provider could cause a null pointer crash during the ExecutorStart phase.

None

Real-time materialized views

Fixed an issue where incremental materialized view refresh produced incorrect data for projection columns containing time functions in the full mutable + group (window function) scenario.

None

Optimizer

• Strengthened equivalence class matching to enhance index join capability for outer joins.

• Fixed an issue where correlated subqueries with aggregation reported errors due to subquery return value type issues.

None

High availability

Fixed a SIGSEGV crash on partitioned tables when = ANY(ARRAY(subquery)) references a non-existent column.

None

Category

Feature

Description

References

Bug fixes

Vector engine

Fixed an issue where the mapping table was incorrectly written in the integrated RAG service.

None

Security

• Fixed a SQL injection vulnerability in the GraphRAG plug-in.

• Fixed security vulnerabilities related to pgml model deserialization, file access, and external dataset loading.

None

Category

Feature

Description

References

New features

Vector engine

• Nova vector index now supports the sq8vector native type.

• Nova now supports the score filter feature.

N/A

Optimizer

• ORCA supports filter condition reordering, and the planner sorting algorithm is optimized.

• Single-column interval merging optimization for filter conditions is added (ORCA and planner).

N/A

Real-time materialized view

Real-time materialized views now support CTE.

N/A

Bug fixes

Stability

• Fixed the NULL handling issue in the fixnumeric stddev_samp function implementation.

• Fixed the issue where executing ALTER AM within a transaction in the new version reports an error.

N/A

PostGIS

Fixed the issue where executing ST_GeomFromWKB(bytea, int) in PostGIS 3.3.2 reports an Unknown WKB type error.

N/A

Long-term memory

Fixed the issue where delete_all in the adbpg_llm_memory plugin recreates the memory table.

N/A

Category

Feature

Description

References

New feature

Real-time materialized view

Real-time materialized views now support CTE.

N/A

Bug fixes

Optimizer

• Fixed the issue where the planner optimizer may miss the CTE producer when a shared CTE is included in a subplan.

• Fixed the coredump issue caused by two-stage window aggregation.

N/A

PostGIS

Fixed the issue where executing ST_GeomFromWKB(bytea, int) in PostGIS 3.3.2 reports an Unknown WKB type error.

N/A

Long-term memory

Fixed the issue where delete_all in the adbpg_llm_memory plugin recreates the memory table.

N/A

Category

Feature

Description

References

New features

vector engine

• Supports the RaBitQ quantization algorithm based on the Novam index.

• Supports Principal Component Analysis (PCA) for dimensionality reduction.

• Supports half-precision (FP16) storage for Nova indexes.

• Optimizes multiple performance issues in the Novam index.

• Adds a write back-pressure mechanism for Nova indexes.

None

monitoring and logging

• Adds parameter information to logs for interrupted statements to simplify issue diagnosis.

• Adds monitoring metrics for query and write response times (RTs) and failure rates.

None

Enhancements

vector engine

• Improves the performance of table lookups for Nova indexes, which reduces end-to-end query latency.

• Optimizes the user-defined function (UDF) that displays delta information for Nova indexes, providing clearer index status details.

None

GraphRAG

Upgrades the GraphRAG component to version 1.6.

None

Bug fixes

vector engine

• Fixed an issue where the graph quality of the Nova index degraded, which improved the search recall rate.

• Fixed a core dump issue caused by the relation cache (relcache) being released after a write-triggered asynchronous analyze or a DDL operation.

• Fixed a memory core dump issue triggered during Nova index searches.

• Fixed a use-after-free crash that occurred when creating Nova indexes multiple times within the same connection.

None

storage engine

Fixed an input key out of range error in the user-defined storage (ussr) engine.

None

scaling

• Fixed an issue where an unexpected context id... error could occur during transaction commits after a single-replica instance is scaled out.

• Fixed an issue that prevented segment scale-in for AnalyticDB for PostgreSQL instances in mirror-independent deployment mode.

None

optimizer

• Fixed an issue in EXPLAIN ANALYZE where init_tuplestore_state was incorrectly accessed in a ShareInputScan operation within an InitPlan.

• Fixed a core dump in the ORCA optimizer caused by a conversion error when handling queries with internal Common Table Expressions (CTEs).

None

high availability

Fixed an issue where master node queries would hang and cause 100% CPU usage when a segment machine was down.

None

executor

Fixed an issue where hitting the disk spill file limit during execution left a lightweight lock, preventing other processes from acquiring it.

None

Category

Feature

Description

References

Enhancements

GraphRAG

Upgrades the GraphRAG component to version 1.5.

None

Bug fixes

executor

Fixed an issue where hitting the disk spill file limit during execution left a lightweight lock, preventing other processes from acquiring it.

None

high availability

Fixed an issue where master node queries would hang and cause 100% CPU usage when a segment machine was down.

None

CDC

Fixed an issue where segment repair caused slot loss when a standby instance was down.

None

pgsearch

Fixed an issue where a dbsize error on a partitioned root table prevented Data Management Service (DMS) from opening the corresponding schema.

None

optimizer

Fixed an error in Multi-Range Table Function (MultiRTF) caused by segment inconsistencies when an append operation was used with multiple one-time filters.

None

real-time materialized view

Fixed an index rebuild issue that occurred during a readable refresh of a real-time materialized view.

None

storage engine

• Fixed an issue that caused unnecessary rewrites of Append-Optimized (AO) tables when non-storage reloptions were modified.

• Fixed an input key out of range error in the user-defined storage (ussr) engine.

None

Category

Feature

Description

References

Bug fixes

Security fix

Fixed the CVE-2026-2006 vulnerability by improving how the SUBSTRING function handles multi-byte characters to prevent buffer overflows and errors from invalid characters.

None

Category

Feature

Description

References

New features

long-term memory

Adds an infer interface for memory import, allowing you to import data directly without extraction by a large language model (LLM).

None

GraphRAG

Supports importing custom chunk and embedding values.

None

Enhancements

high availability

• Standby instances are now marked as down more quickly to prevent user write transactions from hanging.

• Adds support for re-entrant incremental repairs.

None

Bug fixes

CDC

Fixed stability issues with CDC.

None

executor

Fixed stability issues with the executor.

None

connection pool

Fixed an issue where pgbouncer would not restart after an upgrade.

None

storage engine

Fixed stability issues with the beam storage engine.

None

Vector retrieval

Fixed stability issues with pgsearch.

None

scaling

Fixed an access error after a scale-in operation, which was caused by inconsistent metadata for master-only tables used by multiple plug-ins across different nodes.

None

real-time materialized view

Fixed stability issues with real-time materialized views.

None

optimizer

Fixed stability issues with the optimizer.

None

Category

Feature

Description

References

New features

vector engine

Supports configuring the number of flush workers and the total available CPU cores.

None

Bug fixes

high availability

• Fixed an issue that caused prolonged service interruptions when a standby instance went down.

• Fixed issues related to re-entrant incremental repairs.

None

optimizer

Fixed multiple stability issues with the ORCA optimizer.

None

CDC

Fixed multiple stability issues with CDC.

None

real-time materialized view

Fixed multiple stability issues with real-time materialized views. For example, fixed an issue where the rebuild step could not be skipped when you use REPLACE with an identical definition that specified column names.

None

connection pool

Fixed an issue where pgbouncer was incorrectly launched during an upgrade.

None

Category

Feature

Description

References

Enhancements

storage engine

Supports modifying the default access method (am) for temporary tables.

None

Category

Feature

Description

References

New features

vector engine

Major update: The new Nova vector engine is now generally available. This feature is not supported on ARM-based platforms.

Nova vector search engine

executor

Major update: Upgraded the laser parallel executor.

None

logical backup

Supports Database Backup Service (DBS) storage pool interfaces.

None

long-term memory

The long-term memory feature now supports automatic memory classification and custom classification labels. It also includes an infer interface that allows you to import memory directly without extraction by an LLM.

Advanced features of long-term memory for AnalyticDB for PostgreSQL

Security

Introduced multiple security enhancements.

None

tiered storage

The access paths for cold storage, OSS Foreign Data Wrapper (FDW), and OSS EXT now support HTTPS.

None

Enhancements

long-term memory

Improved how historical operations are saved in long-term memory and added a new interface to view historical records.

None

GraphRAG

Supports adjusting the connection pool size.

None

Bug fixes

pgsearch

Fixed multiple stability issues with pgsearch.

None

scaling

Fixed a core dump issue triggered by the expand table command for a real-time materialized view.

None

CDC

Fixed multiple stability issues with CDC.

None

Version upgrade

Fixed an issue where a failed version upgrade caused an instance to enter read-only mode.

None

storage engine

Fixed multiple stability issues with the beam storage engine.

None

real-time materialized view

Fixed multiple stability issues with real-time materialized views.

None

Vacuum

Fixed multiple stability issues related to the VACUUM process.

None

optimizer

Fixed multiple stability issues with the ORCA optimizer.

None

high availability

Fixed multiple stability issues related to high availability.

None