Apache Fluss Roadmap 2026

[wuchong]

This roadmap outlines the strategic direction and key focus areas for the Apache Fluss community in 2026. It highlights both ongoing initiatives and planned features designed to solidify Fluss as the Streaming Storage for Real-Time Analytics & AI.

Disclaimer: This document reflects our current vision and priorities. It is not a strict delivery commitment. We adapt our plans based on community feedback and the evolving data landscape.

---

📢 We Need Your Voice

If you are running Fluss in production or evaluating it for your next-gen architecture, your feedback is critical. Please help us prioritize by letting us know:

• Blockers: What is missing that prevents you from deploying Fluss today?
• Priorities: Which feature below would provide the most immediate value to your team?
• AI & Lakehouse: Which specific integration do you expect the most?

---

🧠 Pillar 1: Real-Time AI & ML

Transforming Fluss into the backbone for real-time AI & ML.

• Real-Time Feature Store: Native support for feature engineering primitives, including:
  • Aggregation Merge Engines for real-time and long-period features
  • Zero-Cost Schema Evolution for safely ML model iteration
  • Point-in-Time Correctness for training data.
• Multimodal Streaming Data: row-based, column-based, vector-based, variant, images, etc.
• Python AI Ecosystem: A high-performance Python SDK designed for AI/ML workflows, enabling seamless integration with PyTorch, Ray, Pandas, and PyArrow for direct access to both real-time streams and historical snapshots.

🌊 Pillar 2: Real-Time Lakehouse

Bridging the gap between Streaming and the Lakehouse with sub-second latency.

• Fluss x Iceberg: See the Fluss x Iceberg 2026 Roadmap
• Union Read Ecosystem: Native Union Read support for Spark, Trino, and StarRocks, etc.
• Performance: Implementation of Deletion Vectors to significantly accelerate Union Reads for updates and deletes.
• Seamless: Capability to define Fluss tables directly on top of existing Lakehouse tables.
• Data Lakes: Integrate with Hudi and Delta Lake.

🚀 Pillar 3: Streaming Analytics (Flink & Spark)

Deep integration to provide a "Single Engine Experience" for unified batch and stream processing.

• Global Secondary index: Enabling fast lookups on any non-primary key columns.
• Delta Join: General Availability of Delta Joins, multi-stream support, left/right/full joins support.
• Real-Time Profiles: Native support for RoaringBitmap aggregation and auto-increment keys to power User/Item Profile use cases.
• Optimizer: Moving from RBO to CBO in Flink SQL by leveraging Fluss table statistics.
• Spark Engine: Full Spark Engine support, including Catalog, Streaming Read/Write, and Batch Read/Write.
• Spark RTM: Integration with Spark 4.1 Structured Streaming Real-Time Mode.

💾 Pillar 4: Storage Engine

• Columnar Streaming: comprehensive support for Filter Pushdown ([Feature] Support filter pushdown for Log Tables #197) and Aggregation Pushdown to minimize data transfer.
• Schema Evolution: Full support for adding new columns, altering existing columns, and dropping columns, rename columns.
• Table Metadata: Support for renaming tables and default values on columns.

☁️ Pillar 5: Cloud-Native Architecture

Simplifying operations and reducing TCO.

• ZooKeeper Removal: Eliminating the ZK dependency to simplify deployment and reduce operational complexity.
• Zero Disks: "Direct write to S3" to make storage nodes diskless and elastic and eliminate inter-zone networking costs.

🔌 Pillar 6: Connectivity & Ingestion

Making it easier to get data in and out of Fluss.

• Log Ingestion Ecosystem: Direct integration with popular agents like Logstash, Vector, Filebeat, Fluentd. Or support Kafka compatibility to allow existing log collectors to write to Fluss without code changes.
• Client SDK Expansion: Rust Client, C++ Client, Python Client.

🛠️ Pillar 7: Operational Excellence

Features to make Day-2 operations seamless.

• Cluster Rebalance: Automated tools to rebalance data distribution across nodes.
• Rescale Tables: Bucket rescaling for Log Tables and Primary Key Tables.
• Coordinator HA: High-available for coordinators (multi-AZ required)
• Geo-Replication: Support cross-cluster data mirroring with something like mirror-maker.

[Jackeyzhe]

Shall we have plan to support rename table in Fluss? Most of the datalakes support it.

[wuchong]

Agree. I added that into the Schema Evolution.

[MehulBatra]

Fluss x Iceberg 2026 Roadmap

Pillar 1: Format & Schema Evolution
Modernizing Iceberg integration with V3 capabilities and seamless schema management.

• Iceberg V3 Support: Deletion Vectors, Puffin files, Row-Level Lineage, Variant datatype, timestamp_ns/timestamptz_ns, Initial/Default Value support
• Schema Evolution: Add column, alter column types, rename column with automatic Fluss → Iceberg sync
• Schema Cleanup: Remove Fluss system columns (__offset, __bucket, __timestamp) from Iceberg tables

Pillar 2: Catalog Ecosystem
Expanding catalog support and documentation.

• Catalog Integration: AWS Glue, Hive, JDBC, Polaris catalog support
• Documentation: All catalogs support Iceberg REST protocol (Fluss already powers Iceberg REST protocol)

Pillar 3: Performance Optimization
Accelerating write/read paths and adaptive tiering.

• Deletion Vectors: Full-path integration (write + read) for 10-100x faster deletes
• Adaptive Tiering: File sizing during tiering based on workload patterns

Pillar 4: Multi-Engine Ecosystem Integration
Enabling unified hot + cold queries across compute engines.

• Batch Union Read: Support across for unified hot + cold queries
• Spark,
• Trino,
• StarRocks

Pillar 5: In-Place Lakehouse Integration
Define Fluss tables on existing Iceberg data without migration.

• Use Iceberg as historical baseline, Fluss as real-time layer

Pillar 6: Maintenance Operations
Lifecycle management for Iceberg metadata and files.

• Snapshot Expiration: Configurable (default: disabled)
• Orphan File Deletion: Configurable (default: disabled)

Pillar 7: Performance Benchmarking
Measuring and validating performance characteristics.

• Tiering throughput, maximum table scalability, Union read performance