In addition to answering questions in natural language, EMR AI Assistant provides 12 expert skills that target specific operations scenarios such as slow queries, execution plans, resource usage, data ingestion, and materialized views. This topic describes the command, target problem, and required input for each skill so that you can pick the right one when an issue occurs.
Expert skills compared with natural language questions
When you ask a question in natural language, EMR AI Assistant must first infer your intent and then decide which data to read. Selecting an expert skill tells the assistant which troubleshooting method to apply. The assistant follows a fixed diagnostic path for that scenario, collects the corresponding evidence, and returns results in a consistent structure.
Select an expert skill when you already know the problem type, such as a failed ingestion job. Describe the symptom in natural language when the problem type is unclear, such as when you only know that queries have become slower, and let the assistant determine the direction.
Invoke an expert skill
Expert skills are invoked with slash commands in the chat window. Both of the following methods work:
Enter a slash (
/) to open the skill menu, and then select a skill from the list.Enter the full command directly, such as
/slow-sql, and then provide the instance, time range, and other details.
After the command, use @ to specify the target instance, and then describe the problem. The complete format is as follows:
/command @instance_name problem description
The skills shown in the menu depend on your account permissions and the instance type. Refer to the menu in the chat window for the skills that are actually available to you.
Expert skill list
The following table describes the command and scope of each expert skill.
Command | Skill | Target problem | Information you provide | Output |
| Slow query analysis | Identifies the slowest SQL statements in a time range and determines whether the bottleneck is the SQL statement, the table design, resource contention, or data skew. | Instance and time range. Include the SQL statement or query ID if you already have one. | Ranked slow SQL statements, bottleneck categories, and optimization suggestions. |
| Query Profile analysis | Analyzes the Query Profile of a single SQL statement and locates the time cost down to the operator level. | Query ID or Query Profile content, together with the SQL statement. | Time cost per stage, bottleneck operators with supporting evidence, and verification steps. |
| Resource diagnostics | Analyzes abnormal CPU, memory, disk, and I/O usage, and determines whether the pressure comes from queries, ingestion, compaction, or a node failure. | Instance, time range, and the symptoms that you observed. | Source of the resource consumption, related jobs, and remediation suggestions. |
| Compaction diagnostics | Locates version accumulation, compaction failures, and the affected tables and partitions. | Alert time, the tables or partitions involved, and the compaction score. | Scope of the accumulation, root cause assessment, and low-risk remediation actions. |
| Table design | Recommends a table type, partitions, buckets, and sort keys for a new table. | Query patterns, write method, data volume, and expected growth. | A table design that covers the table type, partitions, buckets, and sort keys. |
| SQL tuning | Rewrites a known slow SQL statement and optimizes its execution plan. | SQL statement and related table schemas. Include the Query Profile if available. | Rewrite suggestions, expected gains, and the risks of each change. |
| Data ingestion | Troubleshoots ingestion failures, latency, and throughput that falls short of expectations. | Ingestion method, job name, and error message. | Error classification, the point of failure in the pipeline, and fixes. |
| Materialized views | Analyzes refresh failures, inactive states, and queries that fail to hit a materialized view. | Materialized view name, base tables, query statement, and refresh history. | State assessment, refresh issues, and suggestions for improving hit rates. |
| Instance health check | Runs a multi-dimensional health check on an instance. Suitable for periodic checks, pre-change verification, and post-incident review. | Instance and the areas that you want to focus on. | Health score, risk items, and an optimization checklist. |
| Connection diagnostics | Troubleshoots failures when a client cannot connect to an instance. | Client type, connection endpoint, error message, and network environment. | Conclusions grouped by network, authentication, and permissions. |
| Catalog troubleshooting | Troubleshoots data lakehouse catalog configurations and external table access errors. | Catalog type, configuration, and error message. | Conclusions grouped by metadata, storage, and permissions. |
| Error analysis | Analyzes error logs and abnormal events, and assesses the scope of impact. | Error summary, time of occurrence, and instance. | Scope of impact, root cause assessment, and next steps. |
Select a skill based on the symptom
If you are not sure which skill to use, use the following table to work backward from the symptom.
Symptom | Recommended skill |
Users report slower queries, but you do not know which SQL statements are involved. | Slow query analysis ( |
You have identified a slow SQL statement and need to know which step is slow. | Use Query Profile analysis ( |
CPU, memory, or disk usage has increased suddenly. | Resource diagnostics ( |
You received a compaction alert, or queries slowed down after a write operation. | Compaction diagnostics ( |
An ingestion job failed, returned an error, or became slower. | Data ingestion ( |
A materialized view did not refresh as expected, or queries do not hit the materialized view. | Materialized views ( |
A client or BI tool cannot connect to the instance. | Connection diagnostics ( |
Queries against Iceberg, Hive, or other external tables return errors. | Catalog troubleshooting ( |
You are onboarding a new workload and need to confirm the table design. | Table design ( |
You need a health check before or after a change. | Instance health check ( |
Logs contain a large number of errors, but the scope of impact is unclear. | Error analysis ( |
You only know that the business is affected and cannot tell which category the problem belongs to. | Describe the symptom in natural language and let EMR AI Assistant determine the direction. |
Phrase your question
For the same skill, the more complete your input, the more specific the conclusion. Organize your question in the following order.
Enter the skill command, such as
/slow-sql.Use
@to specify the target instance. This prevents context confusion when you manage multiple instances.Provide an explicit time range, such as "from 10:00 to 11:00 on 2026-08-06", instead of "recently" or "just now".
Add key evidence: the SQL statement, query ID, Query Profile, error summary, job name, or alert name.
State the output that you expect. For example, ask the assistant to separate confirmed conclusions from hypotheses that still need verification, or to provide verification steps.
The following table provides sample questions that you can use directly. Replace <INSTANCE_NAME> with the name of your instance.
Scenario | Sample question |
Slow query analysis |
|
Query Profile analysis |
|
Resource diagnostics |
|
Ingestion troubleshooting |
|
Catalog troubleshooting |
|
Instance health check |
|
EMR AI Assistant analyzes the actual runtime data of your instance. The same question asked at different times can return different conclusions. Keep the query ID, time range, and other key details for later review.
Assess whether a conclusion is actionable
A diagnostic output that supports an operations decision usually contains the following items. If an item is missing, ask a follow-up question in the same conversation.
Analysis target: the instance, time range, job, SQL statement, or alert that the analysis covers.
Evidence source: whether the conclusion is based on the Query Profile, monitoring metrics, logs, job execution records, or instance configurations.
Root cause: confirmed conclusions separated from hypotheses that still require verification.
Recommended actions: low-risk actions that you can apply directly, separated from high-risk changes that require approval.
Risk notes: the scope of impact, the rollback method, and the scenarios in which the action does not apply.
Next steps: the expert skill to use next, or the items that require manual inspection.
Limits
EMR AI Assistant returns diagnostic conclusions and recommendations. It does not modify instance configurations, scale instances, or restart instances. You must confirm and apply any change that affects a production environment in accordance with your change management process.
EMR AI Assistant reads operational data within the permissions granted to the current account. An Alibaba Cloud account must grant permissions to RAM users before they can use the assistant. For more information, see EMR AI Assistant quick start.
Diagnostics rely on operational data such as system tables, Query Profiles, monitoring metrics, job execution records, and error summaries. The assistant is not intended to analyze detailed business data in your tables.
Do not enter passwords, AccessKey IDs, AccessKey secrets, tokens, cookies, webhook keys, or unmasked business data in the conversation.
FAQ
Does EMR AI Assistant modify instance configurations automatically?
No. EMR AI Assistant only returns diagnostic conclusions and recommendations. You must apply configuration changes, scaling operations, restarts, and permission adjustments yourself.
Which skill should I use for a specific problem?
If the problem type is clear, select the corresponding skill. For more information, see Select a skill based on the symptom. Otherwise, describe the symptom in natural language, and the assistant determines the direction and suggests next steps.
What do I do if a skill is missing from the menu?
The skills shown in the menu depend on your account permissions and the instance type. Make sure that your account has the required permissions on the instance, and then open the menu again. If the skill is still unavailable, describe the problem in natural language. The assistant follows a similar diagnostic path.
How do I improve diagnostic accuracy?
Provide the instance, time range, symptom, SQL statement or query ID, error summary, and the normal baseline of your workload. The more complete the input, the easier it is for the assistant to reach a verifiable conclusion.
Can I apply the recommendations directly?
You can apply low-risk actions, such as adjusting a query or collecting statistics, after you verify them in a test environment. High-risk actions, such as scaling, parameter changes, and restarts, require manual review against your monitoring data and approval through your change management process.