MaxCompute Mars Sunset & Lakehouse Deprecation Timeline for 2026 - MaxCompute - Alibaba Cloud - MaxCompute

This topic describes MaxCompute product announcements for 2026. It covers new features, syntax and attribute updates, and permission changes designed to improve your project development efficiency.

September 1, 2026 - AI Data Exploration Client Capability Upgrade

Dear MaxCompute users,

The AI Data Explorer client has been upgraded with new capabilities, including MaxFrame Notebook, dbt integration, data dashboards, workplace collaboration, semantic package management, and multimodal dataset management, further enhancing the unified experience for data development and AI applications. The new capabilities are as follows:

MaxFrame Notebook / Python UDF Development
Develop MaxFrame distributed jobs in a notebook interface within the client. You can write and run code cell by cell, preview outputs and target table data in real time, and benefit from pre-execution static checks for common issues such as missing execute() calls and full-table scans. You can also generate complete jobs from natural language prompts using AI. Python UDFs allow you to encapsulate row-level logic that is difficult to express in SQL into reusable functions and call them directly in SQL statements.
dbt and MaxCompute Ecosystem Integration
Integrate dbt data transformation workflows with MaxCompute. Manage SQL logic, dependencies, and lineage in a model-based approach, and edit, run, and review compilation results within the client.
Data Dashboard
Persist SQL query results as dashboards with multiple chart types, cross-chart interactive filtering, parameterized queries, and drill-down capabilities. Dashboards can also be generated from query results with one click using AI.
Workplace Collaboration
Connect to workplace collaboration tools such as DingTalk to share query results, reports, and dashboards directly to collaboration platforms, bringing data insights into team workflows.
Semantic Package Management and Application
Consolidate business metric definitions, indicator specifications, and table relationships into semantic packages for AI reuse, ensuring that natural language query results align with actual business definitions.
Multimodal Dataset Management
Manage unstructured data such as images, audio, and video in tables using the Blob column type. Upload, preview, and batch-process multimodal data, sharing a single data foundation for both structured analysis and multimodal processing.

If you have any questions or need assistance, please contact us by submitting a ticket.

Thank you for your understanding and support of MaxCompute!

August 17, 2026 - PyODPS discontinued maintenance of `odps.apis.storage_api_v2` starting from 0.13.1

Dear MaxCompute users,

Starting from PyODPS version 0.13.1, the odps.apis.storage_api_v2 module is officially discontinued. This module will no longer receive feature updates, bug fixes, or compatibility adaptations, and will be removed in a future version.

Impact scope
This only affects code that directly imports odps.apis.storage_api_v2. Released versions 0.13.1 and earlier can still be installed and run normally, but no further fixes will be provided for any issues encountered. Once you upgrade to a version that removes this module, the affected code will fail with an ImportError and cannot start.
Recommended actions
Please check your production jobs and internal tools for any references to this module and include them in your deprecation plan. Before completing the migration, we recommend pinning your PyODPS version (for example, specifying pyodps<=0.13.1 in your dependencies) to prevent job disruptions caused by upgrades.

If you have any questions or need assistance, contact us by submitting a ticket.

Thank you for your understanding and support of MaxCompute!

July 22, 2026 - MaxCompute Graph End of Maintenance

Dear MaxCompute users,

After a comprehensive review of feature usage and product roadmap, MaxCompute Graph will enter end of maintenance on July 22, 2026 (Wednesday, Beijing Time). Starting from this date, MaxCompute Graph will no longer receive new feature development, bug fixes, or routine maintenance support.

Existing Graph jobs will continue to run during the transition period, but we recommend migrating to alternative solutions as soon as possible. Related documentation will remain available for reference.

If you have any questions or need assistance, submit a ticket.

Thank you for your understanding and continued support of MaxCompute.

July 13, 2026 - MaxCompute Append Delta Table - Hash Cluster Release

Dear MaxCompute users,

MaxCompute now supports Hash Cluster in the Append Delta Table format, combining incremental data processing capabilities with improved query performance.

This feature is suitable for the following scenarios:

Equality filter queries: Perform point lookups or equality filtering on specific columns to reduce the amount of data scanned.
Equality JOIN / GROUP BY: Join or aggregate multiple tables on the same key to reduce Shuffle overhead.

For more information, see Append delta table - Hash cluster (Invitation only).

If you have any questions or need assistance, submit a ticket.

Thank you for your understanding and continued support of MaxCompute.

June 30, 2026 - Access Behavior Change for SQL UDF, Spark, and MaxFrame Jobs to OSS Intranet

Dear MaxCompute users,

Starting June 30, 2026 (Tuesday, Beijing Time), MaxCompute will block SQL UDF, Spark, and MaxFrame jobs from directly accessing OSS intranet addresses. If you need to access OSS intranet, configure access using the following solution. Failure to configure this may result in failed access to OSS intranet addresses. To access OSS VPC addresses, follow these steps:

Create a network connection. For more information, see Create the network connection.

Configure the following parameters:

-- For SQL UDF:
SET odps.session.networklink=networklink_name;

-- For MaxCompute Spark:
spark.hadoop.odps.cupid.eni.enable = true
# Format: region:vpcid. vpcid is the target VPC ID used when setting up the dedicated connection.
spark.hadoop.odps.cupid.eni.info = regionid:vpc-**********

-- For MaxFrame,Add the following parameter in the UDF
@with_network_options(vpc_network_link="networklink_name")
 def xxx():
     pass

Once the network connection is created and the networklink parameter is configured, you can access OSS intranet addresses.

If you have any questions or need assistance, submit a ticket.

Thank you for your understanding and continued support of MaxCompute.

June 17, 2026 - Tunnel Java SDK Version Upgrade Recommendation

Dear MaxCompute users,

To improve upload link stability and exception recovery, the latest Tunnel Java SDK includes exception detection and fast reconnection optimizations for server-side failover scenarios. Some projects are using earlier SDK versions that are susceptible to the following risks. Evaluate the impact and upgrade as soon as possible.

Background
When using Stream Tunnel to upload data, if a server-side failover event occurs (including but not limited to unexpected node restarts or unplanned shutdowns), earlier SDK versions lack the ability to quickly detect such exceptions. This can cause prolonged 502 errors on the client side, affecting upload job stability. The latest SDK version includes targeted optimizations for this scenario to shorten the duration of 502 errors and reduce the impact on upstream services.
Impact
All users who use Tunnel Java SDK for Stream Tunnel uploads, including but not limited to those who directly integrate the Java SDK or write to MaxCompute through Flink or other products.
Recommendations
Upgrade to the following version baselines based on your setup:
- Tunnel Java SDK: upgrade to version 0.51 or later.
- Flink: upgrade to VVR 11.x or later.

Validate job compatibility in a staging environment before switching to production to ensure business continuity.

If you have any questions or need assistance during the upgrade, submit a ticket.

Thank you for your continued support of MaxCompute.

June 10, 2026 - PyODPS 0.13.0 Release Changes

Dear MaxCompute users,

PyODPS 0.13.0 and later versions include the following changes. Please evaluate the impact on your workloads and prepare for migration accordingly.

1. End of Python 2.7 Support

PyODPS 0.13.0 and later versions no longer support the Python 2.7 runtime. If you are currently using PyODPS with Python 2.7, please evaluate your workloads and migrate to Python 3 at your earliest convenience.

2. Removal of Mars Support

As the MaxCompute Mars framework has been decommissioned, PyODPS 0.13.0 and later versions no longer include Mars-related functionality. If you require distributed DataFrame or Tensor computing capabilities, consider migrating to MaxFrame, the MaxCompute distributed AI computing framework with a native Python development ecosystem.

Migration Recommendations

New projects: Use Python 3.7 or later with PyODPS 0.13.0 or later. Do not use Python 2.7 for new projects.
Existing Python 2.7 workloads: Pin your dependency to pyodps==0.12.6 and plan your Python 3 migration at a suitable business window.

Note for DataWorks Users

In future DataWorks platform upgrades, the Python 2.7 runtime will continue to use the fixed PyODPS 0.12.6 version for an extended period, ensuring that existing tasks run without disruption. No manual adaptation is required.

If you have any questions or need assistance, submit a ticket.

Thank you for your continued support of PyODPS and the MaxCompute ecosystem.

June 9, 2026 - MaxCompute Agentic ecosystem tools launched

Dear MaxCompute users,

MaxCompute officially launches the Agentic ecosystem toolkit, which includes five components: AI Data Explorer, MCP Server (MCMCP), MaxCompute Skills, CatalogAPI SDK, and MaxCompute CLI. Together, they cover the full spectrum from desktop interactive analytics to AI Agent automated integration.

1. AI Data Explorer

A desktop data development and analytics tool that integrates SQL development, Catalog browsing, visual analytics, and AI-powered data exploration. Key features include:

SQL development environment: syntax highlighting, intelligent auto-completion, formatting, multi-file management, and query history.
Data visualization: 12 chart types, pivot tables, and data quality analysis.
AI data exploration: conversation mode and Agent mode for natural-language-driven data analysis.
Data management: CSV, Excel, and Parquet import/export, Blob file browsing, and WebDAV mounting.
Supported on macOS 12.3+ (Apple Silicon) and Windows 10+ (64-bit).

If you want to query data using natural language and SQL, and frequently import or export data, the AI Data Explorer is recommended as an all-in-one SQL IDE, Data Agent, and multimodal data management tool.

For more information, see MaxCompute AI data exploration.

2. MCP Server (MCMCP)

An AI Agent data service access layer built on the Model Context Protocol (MCP). It enables agents to directly perform large-scale data analytics and intelligent O&M. Core capabilities include:

Catalog metadata browsing and natural language search.
Table management: creation, data insertion, and metadata updates.
Read-only SQL analytics: cost estimation, asynchronous execution, and result retrieval, with server-side read-only enforcement.
Multiple authentication methods: AccessKey pair, STS, Credentials URI, ECS RAM Role, and default credential chain.
Compatible with OpenClaw, DataWorks Agent, Qwen Code, QoderWork, MaxAgent, and other Agent clients.

If you want to use your own desktop Agent to access MaxCompute for data analytics and O&M, MCMCP combined with scenario-specific Skills is recommended.

For more information, see MaxCompute MCP service.

3. MaxCompute Skills

A general-purpose skill pack for AI Agents that provides semantic capabilities specific to the MaxCompute domain. It includes:

Information Schema semantic pack: designed for system-level O&M, with support for storage pressure diagnostics, cost analysis, job failure analysis, permission auditing, and hot table monitoring.
Common command sets and development templates to lower the barrier for Agent-driven MaxCompute operations.

If you want to use your own desktop Agent to access MaxCompute for data analytics and O&M, MCMCP combined with scenario-specific Skills is recommended.

For more information, see MaxCompute MCP service.

4. CatalogAPI SDK

A RESTful SDK (current version: v0.2.2) for MaxCompute data catalog management. Available for Java and Python, it covers the following metadata resource management capabilities:

Table management: create, query, update, and delete internal and external tables.
Connection management: configure connections for external data sources such as OSS and OTS.
Permissions and data security: fine-grained access control through roles and policies, column-level permissions with taxonomy policy tags, and dynamic data masking with DataPolicy.
Metadata crawling: automatic discovery and crawling of external data source metadata through DataScan.
Model management: multi-version management of machine learning model metadata.

For more information, see CatalogAPI SDK User Guide.

5. MaxCompute CLI

A command-line tool (command name: maxc) invoked through the Alibaba Cloud CLI as aliyun maxc. It is designed for script automation and AI Agent integration:

Structured JSON output for easy programmatic parsing.
Supported on Linux, macOS, and Windows.
Distributed through the Alibaba Cloud CLI. Install or update the Alibaba Cloud CLI to get started.

If you want to use an Agent to schedule and execute data jobs, the MaxCompute CLI is recommended.

For more information, see MaxCompute CLI.

If you have any questions or need assistance, submit a ticket.

Thank you for your continued support of MaxCompute.

April 29, 2026 - Behavior change for MaxFrame groupby apply_chunk

Dear MaxCompute users,

Starting from MaxFrame SDK 2.6, the behavior of the DataFrameGroupBy.mf.apply_chunk method is changing. This update simplifies the index parameter and unifies behavior across the DPE and MaxCompute SQL engines. The new behavior is controlled by the prepend_index_group_keys parameter, which defaults to False for backward compatibility but will switch to True in a future release.

Changes:
- Simplified index parameter: The index parameter now represents only the index type of the user function output and no longer needs to include group keys. The final index composition depends on the group_keys and as_index parameters of the groupby function.
- Unified engine behavior: In both the DPE and MaxCompute SQL engines, the input DataFrame passed to user-defined functions no longer includes group keys by default, eliminating the previous behavioral inconsistency between the two engines.
For detailed migration steps and code examples, see Release notes.
Scope of impact:
MaxFrame users who use the groupby.mf.apply_chunk method. If you do not use this method, this change does not affect your workloads.
Recommended actions:
- Add prepend_index_group_keys=True to all groupby.mf.apply_chunk calls to opt in to the new behavior.
- If you specify the index parameter, remove group keys from it.
- If you use the DPE engine and your user-defined function references group columns, update your code to explicitly retain those columns.

If you have any questions or need assistance during migration, submit a ticket to contact us.

Thank you for your continued support of MaxCompute.

March 30, 2026 - Billing changes for external tables and Data Lakehouse

Dear MaxCompute customer,

To align with the evolution of the Data Lakehouse architecture, MaxCompute will adjust the pay-as-you-go billing model for queries on external tables and Data Lakehouse analytics, effective Tuesday, June 30, 2026 (UTC+8). This adjustment is designed to cover a wider range of data sources and more accurately reflect the true cost of computation.

Key changes:

Billing basis change: The billing metric for input data will change from physical size to logical size. The logical size is the uncompressed raw data, typically 3 to 7 times larger than the physical size. The unit price remains CNY 0.03 per GB, but the calculation now includes a computational complexity factor. The new billing formula is: Logical data volume × Complexity factor × CNY 0.03/GB.
Expanded billing scope: In addition to OSS and OTS, read operations from the following data sources will now be billed: Hologres, RDS, HBase, Lindorm, Hive, HDFS, and Object Tables. The LOAD command will remain free of charge.
Subscription policy: For subscription plans, the cost remains unchanged. These jobs will consume your resource group quota, and no additional external computing fees will be charged.

If you are concerned about potential cost increases, we recommend the following actions:

Migrate your compute jobs to run within the data warehouse or on the data lake.
Recommended: Switch from the pay-as-you-go model to a subscription plan. This is the fastest transition method and allows you to lock in costs.

If you have any questions or need assistance, please submit a ticket.

Thank you for your continued support of MaxCompute.

March 5, 2026 - MaxFrame SDK upgraded to v2.5.0+

Dear MaxCompute user,

To deliver a more stable and feature-rich MaxFrame experience, the MaxFrame SDK has been upgraded to version 2.5.0. This upgrade improves system stability and expands functional capabilities. However, because it includes updates to underlying protocols, compatibility issues may arise in specific scenarios, causing MaxFrame jobs to fail or produce errors.

To avoid disrupting your business continuity and data output efficiency, upgrade your MaxFrame SDK to version 2.5.0 or later as soon as possible.

Before switching to the new version, validate the compatibility of your existing jobs in a staging environment. If a job fails or produces an error, check the job logs to confirm whether the issue stems from protocol incompatibility, and contact technical support immediately.

If you have any questions or need assistance during the upgrade or usage process, submit a ticket.