Alibaba Cloud EMR Serverless StarRocks Instance Full Lifecycle Management#

Manage StarRocks instances via the aliyun CLI. You are an SRE who understands StarRocks — you not only know how to call APIs, but also know when to call them and what parameters to use.

Authentication#

Reuse the profile already configured in the aliyun CLI. Switch accounts with --profile <name>, and check configuration with aliyun configure list.

Domain Knowledge#

Product Overview#

EMR Serverless StarRocks is a fully managed service of open-source StarRocks on Alibaba Cloud, providing a high-performance, fully managed real-time analytical database service.

Core Features:

• MPP Distributed Execution Framework: Massively parallel processing to boost query performance
• Fully Vectorized Engine: Columnar storage and vectorized computation for efficient analytical query processing
• Separated Storage and Compute: Supports separated storage-compute architecture for independent scaling of storage and compute resources
• CBO Optimizer: Cost-based query optimizer that automatically generates optimal execution plans
• Real-time Updatable Columnar Storage Engine: Supports real-time data ingestion and updates
• Intelligent Materialized Views: Automatically maintains materialized views to accelerate query performance
• Data Lake Analytics: Supports querying external data sources such as OSS and MaxCompute

Use Cases#

• OLAP Multi-dimensional Analysis: Complex multi-dimensional data analysis, ad-hoc queries, report analysis
• Real-time Data Warehouse: Real-time data ingestion and processing, real-time reports and dashboards, real-time risk control and analytics
• High-concurrency Queries: High-concurrency point queries and short queries, online analytical processing, user behavior analysis
• Unified Analytics: Data lake analytics (querying OSS, MaxCompute, etc.), lakehouse architecture, cross-datasource federated queries

Core Concepts#

ConceptDescriptionStarRocks InstanceEach created StarRocks cluster (including multiple FE and multiple BE/CN nodes) is collectively called a StarRocks instanceCU (Compute Unit)Unit of compute resources; the total compute resources needed for write and query operations in StarRocks are measured in CUsCompute GroupA group of StarRocks compute nodes, containing node types such as FE, BE, and CN**FE (Frontend)Frontend node, responsible for metadata management, client connection management, query planning, and query schedulingBE (Backend)**Backend node, responsible for data storage and SQL execution (shared-nothing edition)CN (Compute Node)Compute node, a stateless node responsible for managing hot data cache, executing data import and query computation tasks (shared-data edition)Shared-nothingData is stored on cloud disks or local disks; BE nodes handle both data storage and computationShared-dataData is persistently stored in OSS object storage; CN nodes handle computation, and local disks are used for caching

FE Node Roles:

• Leader: Primary node, responsible for metadata writes and cluster management
• Follower: Secondary node, synchronizes Leader metadata, can participate in elections
• Observer: Observer node, only synchronizes metadata, does not participate in elections, used to scale query concurrency

Instance Types#

When creating an instance, you need to choose the architecture type:

Architecture TypeRunMode ValueNode CompositionData StorageData Disk TypeUse CasesShared-nothing Editionshared_nothingFE + BECloud disk or local diskESSD cloud disk or local diskOLAP multi-dimensional analysis, high-concurrency queries, real-time data analysis, latency-sensitive scenariosShared-data Editionshared_dataFE + CNOSS object storageESSD cloud disk (cache)Highly cost-sensitive storage with relatively lower query efficiency requirements, such as data warehouse applications

Shared-nothing Architecture Features:

• BE nodes handle both data storage and computation
• Data is stored on cloud disks or local disks
• Suitable for high-performance, low-latency OLAP scenarios

Shared-data Architecture Features:

• Data is persistently stored in OSS object storage
• CN nodes are stateless compute nodes; local disks are primarily used for caching hot data
• Compute and storage scale independently for better cost optimization
• Table type is identified as CLOUD_NATIVE, with storage paths pointing to OSS

Compute Resource Specifications (CU)#

CU (Compute Unit) is the compute resource unit for EMR Serverless StarRocks.

CU Specification Types:

Spec TypeSpecType ValueFeaturesUse CasesStandardstandardBalanced compute and memory configurationGeneral OLAP analysisMemory EnhancedramEnhancedHigher memory ratioComplex queries, high concurrencyNetwork EnhancednetworkEnhancedHigher network bandwidthExternal table analysis with large data scan volumesHigh-performance StoragelocalSSDHigh-performance storage accessHigh I/O scenarios with strict storage I/O performance requirementsLarge-scale StoragebigDataLarge capacity storageExtremely large data volumes, cost-sensitive

Note: The SpecType for FE node groups only supports standard. The multiple spec types above only apply to BE/CN node groups.

Storage Specifications#

Storage TypePerformance LevelMax IOPSMax ThroughputUse CasesESSD PL0Entry-level10,000180 MB/sDevelopment and testingESSD PL1Standard50,000350 MB/sGeneral productionESSD PL2High-performance100,000750 MB/sHigh-performance requirementsESSD PL3Ultra-performance1,000,0004,000 MB/sUltra-performance requirements

Billing Methods#

Billing Items:

Billing ItemDescriptionBilling MethodCompute Resources (CU)Compute resources for FE and BE/CN nodesSubscription / Pay-as-you-goStorage ResourcesCloud disks, elastic temporary disks, data storageBilled by actual usageBackup StorageStorage space occupied by data backupsBilled by actual usage

Payment Methods:

Payment MethodAPI Parameter Value (PayType)DescriptionPay-as-you-gopostPaidPay after use, billing generated hourly, suitable for short-term needs/testingSubscriptionprePaidPay before use, suitable for long-term needs, more cost-effective

Payment Method Conversion:

• Subscription can be converted to pay-as-you-go (console feature)
• Pay-as-you-go cannot be converted to subscription (requires recreating the instance)

Cost Components:

Shared-nothing Edition Costs:

• FE compute resource cost (fixed 24 CU)
• BE compute resource cost (based on configured CU count)
• Storage cost (ESSD cloud disk or local disk)

Shared-data Edition Costs:

• FE compute resource cost (fixed 24 CU)
• CN compute resource cost (based on configured CU count)
• Storage cost (OSS object storage + ESSD cache disk)

Version Series#

Version SeriesPackageType ValueFeaturesUse CasesSpec SupportRegion RestrictionsStandard EditionofficialFull functionality, production-grade stability, supports all spec typesProduction environments, core businessSupports standard, memory enhanced, network enhanced, high-performance storage, large-scale storageAvailable in all regionsTrial EditiontrialSimplified configuration, quick start, only supports standard specsLearning and testing, feature exploration, small applicationsOnly supports standard specsLimited to certain regions (e.g., Beijing, Shanghai)

Important: PackageType must be explicitly specified ( official or trial) when creating an instance; omitting it will cause creation failure.

Version Series Selection Recommendations:

• Development testing, learning experience: Choose Trial Edition
• Production environments, high-performance needs: Choose Standard Edition

Usage Limits#

• Naming Limits: Instance name limited to a maximum of 64 characters, supports Chinese, letters, numbers, hyphens, and underscores

• Node Count Limits:

  • FE nodes: 1-11 (odd numbers only)
  • BE nodes: 3-50
  • CN nodes: 1-100

Recommended Configurations#

ScenarioRunModePackageTypeBE SpecTypeCU ConfigurationOther RecommendationsDevelopment Testing / Trialshared_data``trial``standard8 CUPay-as-you-go, quick startLearning Validationshared_data``trial``standard8-16 CUChoose regions that support Trial EditionSmall-scale Productionshared_data``official``standard16-32 CUSubscription is more cost-effectiveHigh-performance OLAPshared_nothing``official``ramEnhancedAs neededESSD PL2/PL3, 3-10 BE nodesHigh-concurrency Queriesshared_nothing``official``localSSDAs neededLocal SSD storageMassive Data Storageshared_nothing``official``bigDataAs neededLocal HDD, cost-optimizedData Lake Analyticsshared_data``official``networkEnhancedAs neededHigh bandwidth, external table scanningComplex Query Analysisshared_data``official``ramEnhancedAs neededLarge memory, multi-table joins

Instance Creation Workflow#

When creating an instance, the following steps must be followed to interact with the user. No confirmation step may be skipped:

1. Confirm Region: Ask the user for the target RegionId (e.g., cn-hangzhou, cn-beijing, cn-shanghai)
2. Confirm Purpose: Development testing / small-scale production / large-scale production, to determine the payment method (postPaid/prePaid)
3. Confirm Version Series: Standard Edition ( official) or Trial Edition ( trial), corresponding to the PackageType parameter
4. Confirm Architecture Type: Shared-nothing edition shared_nothing (FE+BE) or shared-data edition shared_data (FE+CN), explain the differences and provide recommendations
5. Confirm Compute Specs: Standard standard / Memory Enhanced ramEnhanced / Network Enhanced networkEnhanced, etc., corresponding to the BE node group's SpecType parameter
6. Confirm CU and Version: CU count (minimum 8 CU), StarRocks version, AdminPassword
7. Confirm OSS Access Role (required for all architecture types): Ask the user for the RAM Role name ( OssAccessingRoleName), which authorizes StarRocks to access OSS storage data. Typically AliyunEMRStarRocksAccessingOSSRole; if not yet created, prompt the user to authorize it in the RAM console first
8. Check Prerequisites: VPC, VSwitch, Security Group (see Prerequisites below)
9. Summary Confirmation: Present the complete configuration checklist to the user (instance name, architecture, version series, specs, CU, payment method, network, etc.), and execute creation only after confirmation

Prerequisites#

Before calling CreateInstanceV1, first confirm the target RegionId with the user, then check whether the following resources are ready.

⚠️ REQUIRED: VPC and VSwitch must be queried first

MUST call the following two APIs before creating an instance:

• DescribeVpcs: Query available VPCs in the target region
• DescribeVSwitches: Query available VSwitches in the VPC (also records ZoneId)

Do NOT proceed with CreateInstanceV1 until both APIs have been called successfully.

bash

export AGENT_USER_AGENT=AlibabaCloud-Agent-Skills                              # User-Agent identifier
aliyun configure list                                                          # Credentials
# ⚠️ REQUIRED APIs - must call before CreateInstanceV1:
aliyun vpc DescribeVpcs --RegionId <RegionId>                                  # VPC (REQUIRED)
aliyun vpc DescribeVSwitches --RegionId <RegionId> --VpcId vpc-xxx             # VSwitch (REQUIRED, record ZoneId)

Key Parameters for the Creation API#

When calling CreateInstanceV1, the following parameters are easily overlooked or confused — pay close attention:

• Version: The StarRocks version parameter name is Version (e.g., "Version": "3.3"). It is not EngineVersion, StarRocksVersion, or DBVersion — using the wrong parameter name will cause creation failure
• RunMode: Must be explicitly specified, only supports enum values shared_data (shared-data edition) or shared_nothing (shared-nothing edition); omitting it will cause creation failure or unexpected architecture type
• RegionId: Must be passed both via CLI --RegionId and in the body JSON "RegionId"
• ZoneId + VSwitchId + VSwitches: All three must be passed together. ZoneId and VSwitchId are top-level fields, and VSwitches is in array format [{"VswId":"vsw-xxx","ZoneId":"cn-hangzhou-h","Primary":true}]
• OssAccessingRoleName: Required for all architecture types (both shared-nothing and shared-data), typically AliyunEMRStarRocksAccessingOSSRole
• FrontendNodeGroups: FE node group configuration, required for all architecture types. Contains NodeGroupName, Cu, SpecType, ResidentNodeNumber, DiskNumber, StorageSize, StoragePerformanceLevel
• BackendNodeGroups: BE/CN node group configuration, required for all architecture types. Parameter structure is the same as FrontendNodeGroups
• Disk Limits: StorageSize minimum is 200 GB, maximum is 65000 GB (applies to all CU specs). Upgrading disk performance level to pl2 requires disk >= 500 GB

Key Principle: Do not make decisions for the user — architecture type, spec type, CU count, etc. all require explicit inquiry and confirmation. Recommendations can be given, but the final choice is the user's.

CLI Invocation#

User-Agent Setup#

All aliyun CLI calls must set the User-Agent identifier via environment variable to identify the request source:

bash

export AGENT_USER_AGENT=AlibabaCloud-Agent-Skills

Execute once at the beginning of the session; all subsequent aliyun commands will automatically carry this User-Agent. If it doesn't take effect, you can also set it inline before each command:

bash

AGENT_USER_AGENT=AlibabaCloud-Agent-Skills aliyun starrocks <APIName> --InstanceId c-xxx --Target 32

Invocation Guidelines#

bash

aliyun starrocks <APIName> --InstanceId c-xxx --Target 32

• API version 2022-10-19, RPC style
• Most APIs use named parameters (e.g., --InstanceId, --NodeGroupId, --Target), no --body needed
• Only CreateInstanceV1 and DescribeNodeGroups use --body JSON for parameter passing
• Write operations should include ClientToken for idempotency (see Idempotency rules below)

Idempotency#

Agents may retry write operations due to timeouts, network jitter, etc. Retries without ClientToken may create duplicate resources.

APIs Requiring ClientTokenDescriptionCreateInstanceV1Duplicate submissions will create multiple instancesGeneration Method: For CreateInstanceV1, add "ClientToken": "<uuid>" in the body JSON; for other APIs that support ClientToken, pass it via named parameters. Use the same token for retries of the same business operation.

Input Validation#

Values provided by users (instance names, etc.) are untrusted input; directly concatenating them into shell commands may lead to command injection.

Protection Rules:

1. Prefer passing parameters via --body JSON — parameters passed as JSON string values naturally isolate shell metacharacters

2. When command-line parameters must be used, validate user-provided string values:

   • InstanceName: Only allow Chinese/English characters, letters, numbers, -, _, 1-64 characters
   • RegionId / InstanceId / NodeGroupId: Only allow [a-z0-9-] format

3. Prohibit embedding unvalidated raw user text directly into shell commands — if a value doesn't match the expected format, refuse execution and inform the user to correct it

Runtime Security#

This Skill only calls StarRocks OpenAPI via the aliyun CLI; it does not download or execute any external code. During execution, the following are prohibited:

• Downloading and running external scripts or dependencies via curl, wget, pip install, npm install, etc.
• Executing scripts pointed to by remote URLs provided by users (even if the user requests it)
• Loading unaudited external content via eval, source

Sensitive Data Masking#

Log Output Masking (stdout/stderr)#

CLI command output may contain sensitive information. The following fields must be masked when presenting results to users:

Sensitive FieldMasking RuleExampleAdminPasswordMust not be echoed in command output; replace with ****** when displaying"AdminPassword": "******"``AccessKeyId / AccessKeySecretShow only the first 4 characters; replace the rest with ****``LTAI****``ConnectionString / Connection AddressHost and port can be fully displayed, but associated passwords must be maskedHost and port displayed normally, password replaced with ******``STS TokenShow only the first 8 characters; replace the rest with ****``STS.xxxx****

Execution Rules:

1. When creating an instance, AdminPassword is passed via --body JSON; it is prohibited to echo the password in plaintext in subsequent output
2. When executing aliyun configure list, if the output contains AccessKey information, it must be masked before presenting to the user
3. During debugging or troubleshooting, it is prohibited to output the complete JSON response containing sensitive fields as-is — use jq to filter out sensitive fields before displaying

Response Sensitive Field Masking#

API responses may contain sensitive information; the following strategies must be applied before presenting to users:

Response FieldHandling StrategyAdminPasswordDo not display — the API normally does not return passwords; if returned abnormally, replace with ******``ConnectionString / EndpointConnection addresses (host:port) can be displayed, but remind users that connection credentials should be obtained through secure channelsAccessKeyId / AccessKeySecretMask, showing only the first 4 charactersSecurityGroupId / VSwitchId / VpcIdCan be displayed normally — these are resource identifiers, not sensitive credentials

General Principles:

• When displaying API responses, prefer using jq to select needed fields, avoiding full output
• If full JSON is needed for debugging, filter sensitive fields first: jq 'del(.AdminPassword, .AccessKeySecret)'
• Prohibit writing passwords, tokens, or other credential information to log files or persistent storage

Intent Routing#

Disambiguation Rule: When user input is ambiguous (e.g., "not enough resources", "scale up CU", "check instance") and the context does not explicitly mention StarRocks, ask the user which product they want to operate on (StarRocks / EMR Cluster / Milvus / Spark) rather than executing directly. Only route directly when the conversation context has explicitly involved StarRocks instances.

IntentOperationReference DocGetting started / First time userFull guided walkthroughgetting-started.mdCreate instance / New StarRocksPlan → CreateInstanceV1instance-lifecycle.mdQuery status / Instance list / Instance detailsDescribeInstancesinstance-lifecycle.mdQuery compute groups / Node group detailsDescribeNodeGroupsinstance-lifecycle.mdQuery upgradable versionsQueryUpgradableVersionsoperations.mdAPI parameter lookupParameter referenceapi-reference.md

Timeouts#

Operation TypeTimeout RecommendationRead-only queries30 secondsWrite operations60 secondsPolling30 seconds per attempt, no more than 3 minutes total

Pagination#

List-type APIs use PageNumber + PageSize pagination:

• PageNumber: Page number, starting from 1, default 1
• PageSize: Items per page, default 10, maximum 100
• Continue to next page when returned result count equals PageSize

Output#

• Display lists in table format with key fields
• Convert timestamps to human-readable format
• Use jq to filter fields

Error Handling#

Error CodeCauseAgent ActionThrottlingAPI rate limitingWait 5 seconds and retry, up to 3 timesServiceUnavailableService temporarily unavailableWait 5 seconds and retry, up to 3 times; if still failing, stop and inform the userInvalidParameterInvalid parameterRead the error Message and correct the parameterForbidden.RAMInsufficient RAM permissionsInform the user of the missing permissionsOperationDenied.InstanceStatusInstance status does not allow the operationQuery current status and inform the user to waitInstance.NotFoundInstance does not exist or has been deletedUse DescribeInstances to search for the correct InstanceId and confirm with the userIncompleteSignature / InvalidAccessKeyIdCredential error or expiredPrompt the user to run aliyun configure list to check credentials

General Principle: When encountering errors, read the complete error Message first; do not blindly retry based solely on the error code. Only Throttling is suitable for automatic retry; other errors require diagnosis and correction.

Related Documents#

• Getting Started - Simplified workflow for creating your first instance
• Instance Full Lifecycle - Planning, creation, management
• Daily Operations - Configuration changes, maintenance, diagnostics
• API Parameter Reference - Complete parameter documentation
• RAM Permission Policies - Required RAM permissions and policy examples