E-MapReduce (EMR) Serverless StarRocks is a fully managed StarRocks service on Alibaba Cloud. Create StarRocks instances and manage instances and data from the EMR console — without configuring, operating, or scaling clusters yourself.
What is StarRocks
StarRocks is an analytic database built for fast, real-time, and efficient multi-dimensional data analytics. It uses an MPP architecture with a vectorized execution engine, a cost-based optimizer (CBO), intelligent materialized views, and a real-time updatable columnar storage engine. StarRocks is compatible with the MySQL protocol, so any MySQL client or common BI tool can connect directly. It scales horizontally and provides high availability (HA) and high reliability.
StarRocks loads data from real-time sources or an offline store and supports these data analytics scenarios:
Real-time data warehouses — Synchronize changes from transactional databases in seconds and query up-to-date data
Online analytical processing (OLAP) — Run multi-dimensional reports, self-service dashboards, and ad hoc queries
Data lake analysis — Query data in your data lake without migrating it
Core capabilities
The following capabilities are built into the StarRocks engine.
MPP framework
StarRocks splits every query into physical computing units that use parallel execution across machines, each with dedicated CPU and memory. When you scale out the cluster, single-query performance scales with it.
Vectorized execution engine
The vectorized execution engine optimizes all execution operators, functions, scanning and filtering modules, and import and export modules at the CPU level. It uses single instruction multiple data (SIMD) instructions to process more data per clock cycle. Tests on a standard test set show a 3x to 10x improvement in overall operator performance.
The engine also includes Operation on Encoded Data, which runs join, aggregation, and expression operators directly on encoded strings without decoding. This reduces SQL execution complexity and improves query speed by more than 2x.
Compute-storage separation
Introduced in StarRocks 3.0, the compute-storage separation architecture decouples each compute resource from each storage resource so they can scale independently and reduce storage cost. Compute nodes scale within seconds, eliminating the over-provisioning required when compute and storage must grow together.
The storage layer uses various object storage services with nearly unlimited capacity and is compatible with Hadoop Distributed File System (HDFS). The compute-storage separation architecture retains full feature parity with the compute-storage integration architecture. Data updates, data lake analysis, and materialized view acceleration all work the same way. The performance of data writing and hot data query is almost the same in the two architectures.
Cost-based optimizer
In complex multi-table join queries, the number of valid execution plans grows exponentially with the number of tables, making optimal plan selection NP-hard. The StarRocks CBO uses a cascades-like architecture customized for the vectorized execution engine. It supports:
Common sub-expression reuse and correlated subquery rewriting
Lateral Join and Join Reorder
Distributed join execution policy selection
Low-cardinality dictionary encoding optimization
The CBO supports all 99 TPC-DS SQL statements.
Real-time columnar storage engine
StarRocks stores data in a columnar format. This improves compression ratios, reduces disk I/O and the data volume read, and accelerates queries that read only a subset of columns — the common pattern in OLAP workloads. StarRocks allows you to load data within seconds and provides near-real-time data processing capabilities.
The storage engine guarantees ACID (atomicity, consistency, isolation, and durability) for data imports. Batch imports either succeed or fail atomically, and concurrent transactions benefit from snapshot isolation. The engine also supports partial update and upsert operations. It uses primary key indexes with a Delete-and-Insert mode to avoid sort-and-merge overhead during reads. Secondary indexes handle high-throughput data update scenarios.
Intelligent materialized views
Materialized views in StarRocks work automatically:
Automatic synchronization — When data in a source table changes, the corresponding materialized view detects and applies the update in real time, keeping data consistent
Transparent query rewriting — During query planning, StarRocks detects when a materialized view can accelerate a query and performs query rewrite automatically; no application changes required
Background lifecycle management — Create and delete materialized views without manual intervention; the system handles the operation in the background
ETL replacement — Use materialized views to transform and process data in place, replacing traditional extract, transform, and load (ETL) pipelines and upstream pre-processing
Data lake analysis
StarRocks analyzes local storage data efficiently. Use external catalogs to query data lakes directly — no data migration required. StarRocks supports:
Table formats: Apache Hive, Apache Iceberg, Apache Hudi, and Delta Lake
File formats: Parquet, ORC, and CSV
Storage services: HDFS, Amazon Simple Storage Service (S3), and Object Storage Service (OSS)
In this model, data lakes serve as the single source of truth (SSOT) for BI, AI, ad hoc queries, and reporting workloads. StarRocks handles compute and analysis using its vectorized engine and CBO.
What Serverless StarRocks adds
Running StarRocks yourself means provisioning clusters, planning version upgrades, configuring security, and monitoring the system. EMR Serverless StarRocks eliminates this operational overhead:
No cluster management — Skip cluster sizing, setup, and ongoing tuning
Instance management visualization — Manage instances and run O&M tasks from the EMR console
Visualized monitoring — Built-in monitoring and O&M dashboards
Automatic version upgrades — Major and minor StarRocks versions upgrade automatically
EMR StarRocks Manager:
Security: Provide user and permission management
Diagnostic analysis: Identify slow SQL statements and analyze SQL execution with visual tools
Data management: Browse databases, tables, partitions, shards, and tasks to streamline operations management
AI+OLAP (Beta) — Call large language models with SQL in your OLAP analytics workflow, as described in AI capabilities.
AI capabilities
EMR Serverless StarRocks includes built-in AI Functions that embed large language model capabilities into your OLAP analytics workflow. AI Functions are in Beta. Process data, analyze it, and run AI inference in a single SQL statement, without exporting data to an external system first.
Version requirements
The instance must run version 3.3.20-2.1.1, 3.5.16-2.1.1, or later.
Key benefits
Data stays in StarRocks: Raw data always remains in StarRocks. AI functions encrypt and send only the fields that need processing to the model service, which prevents data replication and cross-system transfers.
Call with standard SQL: Call built-in AI Functions from standard SQL — no new programming language or tool to learn.
Results feed directly into your queries: AI function results take part in subsequent joins, aggregations, and filters immediately, with no second data load.
Lower token costs: Predicate pushdown reduces call volume and caching eliminates duplicate requests, which lowers the cost of the same task by 30% or more.
Rate limiting without O&M: Three built-in protection layers keep your workloads stable, so you do not need to understand rate limiting mechanisms such as RPM and TPM.
Function capabilities
AI Functions are grouped by capability type:
| Capability type | Typical functions | Description |
| Text analysis | ai_sentiment, ai_classify, ai_extract, ai_summarize | Runs sentiment analysis, tag categorization, entity extraction, and content summarization on text. |
| Text processing | ai_translate, ai_fix_grammar, ai_redact | Provides machine translation, grammar correction, and masking of personally identifiable information (PII). |
| Semantic filtering | ai_filter, ai_similarity | Filters data by semantic conditions, or calculates the semantic similarity between two pieces of text. |
| Vectorization | ai_embed, ai_embed_multimodal | Generates text embeddings or a multimodal embedding for content such as images and videos, for retrieval scenarios. |
| AI aggregation | ai_agg, ai_agg_summary | Summarizes multiple rows by dimension with AI, such as summarizing all reviews of a product. |
| Custom models | ai_custom_query, ai_custom_embedding, ai_custom_multimodal_embedding | Calls the custom models that you registered in AI center. |
Supported models
Text generation uses qwen3.6-plus. Text vectorization uses qwen3-vl-embedding or text-embedding-v4. Multimodal embedding generation uses qwen3-vl-embedding or tongyi-embedding-vision-plus. qwen3-vl-embedding is a unified vector space model, so text embeddings compare directly with image and video embeddings across modalities, which suits cross-modal retrieval.
Typical scenarios
Semantic filtering and classification routing, such as filtering complaints out of support tickets and assigning them automatically
Multimodal retrieval and training set construction
AI summaries by dimensions such as customer and product
Sensitive data masking before analysis
Batch content generation and translation
Intelligent O&M
Beyond AI Functions, EMR Serverless StarRocks provides EMR AI Assistant. It connects to your instance and reads system tables, Query Profile (query execution details), and monitoring metrics for performance diagnostics and health inspections, as described in EMR AI Assistant.
The AI Function topic documents function syntax and metric descriptions, and AI Function Best Practices walks through scenario-based usage.
FAQ
What is the default QPS limit for the OSS bucket used by EMR Serverless StarRocks, and how can I monitor it?
The default QPS limit for the OSS bucket is 20,000, raised from 10,000. Dedicated monitoring metrics for StarRocks OSS bucket throttling adjustments are not available yet. As an alternative, configure QPS alerts in Cloud Monitor.
Why does StarRocks FE leader node failover occur? Is it related to Full GC?
FE leader node failover can result from a new leader election that Full GC triggers. When an FE node runs Full GC, a new leader election starts and the FE leader changes. This is normal behavior of the G1 garbage collector, which uses a large heap memory allocation of up to 51 GiB and prioritizes pause times. It does not indicate a system failure.