All Products
Search

This Product

    E-MapReduce:Query profile

    Document Center

    E-MapReduce:Query profile

    Last Updated:Mar 26, 2026

    Analyze queries by using query profile

    A query profile records the execution details of every node involved in a query. Use query profiles in StarRocks Manager to visualize query execution and quickly pinpoint performance bottlenecks.

    Enable query profile

    Set the enable_profile session variable to true:

    SET enable_profile = true;
    To enable query profile globally without affecting query performance, your StarRocks kernel version must be 2.5.13 or later (2.x series) or 3.1.5 or later (3.x series). If your version is earlier, upgrade before enabling; otherwise, query performance may be affected.

    Query profile structure

    A query profile has five hierarchical levels:

    Level Description
    Fragment An execution tree. A query consists of one or more fragments.
    Fragment instance Each fragment can have multiple instances, each running on a different compute node.
    Pipeline An execution chain of connected operators. A fragment instance splits into multiple pipelines.
    Pipeline driver An instance of a pipeline. A pipeline can have multiple pipeline drivers, each running on a separate CPU core to maximize core utilization.
    Operator A computational unit. A pipeline driver consists of multiple operators.

    Analyze a query profile

    StarRocks Manager visualizes query profiles in a tree structure. Tree nodes represent aggregated operators.

    image

    • Click an operator to view its details in the tabs on the right side of the page.

    • If no operator is selected, the right side shows an overview of the query.

    Tip: The system highlights the top three time-consuming operators with progressively darker colors. Start your analysis there — click each highlighted operator to examine its CoreMetrics and identify the bottleneck. Use the Zoom In and Zoom Out buttons, or scroll your mouse wheel, to navigate the profile tree.

    Analysis workflow

    Use this sequence to move from a high-level overview to a specific bottleneck:

    1. Check the execution overview. On the Execution Details tab, look at the time breakdown to understand where the query spent most of its time:

      • CPU-bound: High CumulativeCpuTime relative to ExecutionWallTime

      • I/O-bound: High I/O time or large DiskReadBytes / RemoteReadBytes

      • Network-bound: Large Bytes sent over network

      • Scheduling overhead: High CumulativeWaitTime

    2. Locate the top operators. In the profile tree, the three darkest-highlighted operators are the most expensive. Click each one.

    3. Inspect operator details. In the right-side tabs:

      • CoreMetrics — the most relevant metrics for this operator type

      • NodeMetrics — all metrics for the operator

      • Pipeline — scheduling metrics (usually not the focus)

    Execution overview metrics

    The Execution Details tab displays query-level summary metrics.

    Total

    Metric Description
    Total Total time consumed by the query, including planning, execution, and profiling phases.
    PlannerTotalTime Time spent planning the query, including SQL parsing, analysis, optimization, and execution plan generation.
    ExecutionWallTime Wall-clock time for query execution and backend (BE) processing.
    CollectProfileTime Time spent collecting profile metrics from BE nodes.

    Execution time

    Metric Description
    ExecutionWallTime Total wall-clock time for query execution.
    CumulativeCpuTime Sum of CPU time across all BE nodes. Because this sums across concurrent processes, it exceeds wall-clock time.
    CpuUtilization CumulativeCpuTime divided by ExecutionWallTime. Represents the average number of CPU cores used.
    CumulativeWaitTime Cumulative pipeline wait time, including scheduling wait time (ScheduleTime) and blocking time (PendingTime). Examples: empty input queues, full output queues, unmet dependencies.
    OperatorCumulativeTime Sum of execution times for all operators, covering I/O, network, and computation.
    I/O Cumulative I/O time for all SCAN operators, including local disk reads, remote reads, and cache access.

    I/O

    Metric Description
    RawRowsRead Total rows scanned by all SCAN nodes.
    DiskReadBytes Total compressed data read by all SCAN nodes.
    LocalDiskReadBytes Total compressed data read from local cache by all CONNECTOR_SCAN nodes. Applies only to shared-data instances.
    RemoteReadBytes Total compressed data read from Object Storage Service (OSS) by all CONNECTOR_SCAN nodes. Applies only to shared-data instances.
    ResultRows Total output records generated by all SCAN nodes.
    ResultBytes Total data read by all SCAN nodes.

    Network

    Metric Description
    Bytes sent over network Total bytes transmitted by all Exchange nodes (BytesSent).

    Operator metrics

    Query-level metrics

    Summary metrics

    Metric Description
    Total Total time consumed by the query, including planning, execution, and profiling phases.
    QueryCpuCost Cumulative CPU time used by the query. Sums CPU time across all concurrent processes, so it exceeds actual wall-clock time.
    QueryMemCost Total memory consumed by the query.
    Variables Variables used in the query.

    Pipeline-level metrics

    Metric Description
    ActiveTime Time during which a driver runs.
    DriverTotalTime Total time consumed by a driver.
    PendingTime Time a driver waits when input or prerequisites are not met.

    General metrics

    Metric Description
    OperatorTotalTime Total time consumed by the operator.
    PushRowNum Cumulative output rows generated by the operator.
    PullRowNum Cumulative input rows processed by the operator.
    PullChunkNum Cumulative input chunks processed by the operator.
    PushChunkNum Cumulative output chunks generated by the operator.
    PeakMemoryUsage Maximum memory used by the operator.

    OLAP Scan operator

    Metric Description
    Table Name of the table.
    ScanTime Cumulative scan time. Scan operations run in an asynchronous I/O thread pool.
    TabletCount Number of tablets.
    PushdownPredicates Number of predicates pushed down to storage.
    BytesRead Size of data read.
    CompressedBytesRead Size of compressed data read.
    IOTime Cumulative I/O time.
    BitmapIndexFilterRows Rows filtered by a bitmap index.
    BloomFilterFilterRows Rows filtered by a Bloom filter.
    SegmentRuntimeZoneMapFilterRows Rows filtered by a runtime zone map.
    SegmentZoneMapFilterRows Rows filtered by a zone map.
    ShortKeyFilterRows Rows filtered by a short key.
    ZoneMapIndexFilterRows Rows filtered by a zone map index.

    Connector Scan operator

    The following metrics apply only to instances that use storage-compute separation.

    Metric Description
    CompressedBytesReadLocalDisk Compressed data read from the local cache of a compute node.
    CompressedBytesReadRemote Total compressed data read from OSS.
    IOTimeLocalDisk I/O time to read data from the local cache.
    IOTimeRemote I/O time to read data from OSS.

    Exchange operator — Sink

    Metric Description
    PartType Data distribution mode. Valid values: UNPARTITIONED, RANDOM, HASH_PARTITIONED, BUCKET_SHUFFLE_HASH_PARTITIONED.
    BytesSent Size of data sent.
    OverallThroughput Throughput rate.
    NetworkTime Time to transmit a data packet, excluding post-reception processing time.
    WaitTime Wait time caused by a full sender queue.
    NetworkBandwidth Network bandwidth.

    Exchange operator — Source

    Metric Description
    SenderWaitLockTime Time spent waiting for a lock.
    BytesReceived Size of data received.
    DecompressChunkTime Time spent decompressing data.
    DeserializeChunkTime Time spent deserializing data.
    SenderTotalTime Total time spent sending data.

    Aggregate operator

    Metric Description
    GroupingKeys The GROUP BY columns.
    AggregateFunctions The aggregate functions.
    AggComputeTime Time spent computing aggregate functions.
    ExprComputeTime Time spent computing expressions.
    HashTableSize Size of the hash table.

    Join operator — Probe

    Metric Description
    DistributionMode Data distribution mode.
    JoinType Join type.
    OtherJoinConjunctEvaluateTime Time spent evaluating other join conjuncts.
    ProbeConjunctEvaluateTime Time spent evaluating probe conjuncts.
    SearchHashTableTime Time spent querying the hash table.
    WhereConjunctEvaluateTime Time spent evaluating WHERE conjuncts.

    Join operator — Build

    Metric Description
    JoinPredicates The join predicates.
    JoinType Join type.
    BuildBuckets Number of buckets in the hash table.
    BuildHashTableTime Time spent building the hash table.
    RuntimeFilterBuildTime Time spent building runtime filters.
    RuntimeFilterNum Number of runtime filters.
    DistributionMode Data distribution mode.

    Window Function operator

    Metric Description
    ComputeTime Time spent computing window functions.
    PartitionKeys The partition key column.
    AggregateFunctions The aggregate functions.

    Sort operator

    Metric Description
    SortKeys The sort key.
    SortType Sorting scope: all results, or only the top N results.
    MergingTime Time spent merging data.
    SortingTime Time spent sorting data.

    Table Function operator

    Metric Description
    TableFunctionExecTime Time spent computing table functions.
    TableFunctionExecCount Number of times the table function is executed.

    Project operator

    Metric Description
    ExprComputeTime Time spent computing expressions.
    CommonSubExprComputeTime Time spent computing common sub-expressions.

    Local Exchange operator

    Metric Description
    Type Local Exchange type. Valid values: Passthrough, Partition, Broadcast.
    ShuffleNum Number of shuffles. Applies only when Type is Partition.