KqpLoad

Runs general performance testing for the YDB cluster by loading all components via the Query Processor layer. The load is similar to that from the workload YDB CLI subcommand, but it is generated from within the cluster.

You can run two types of load:

  • Stock: Simulates a warehouse of an online store: creates multi-product orders, gets a list of orders per customer.
  • Key-value: Uses the DB as a key-value store.

Before this test, the necessary tables are created. After it's completed, they are deleted.

Actor parameters

The basic actor parameters are described below. For the full list of parameters, see the load_test.proto file in the YDB Git repository.

Parameter Description
DurationSeconds Load duration in seconds.
WindowDuration Statistics aggregation window duration.
WorkingDir Path to the directory to create test tables in.
NumOfSessions The number of parallel threads creating the load. Each thread writes data to its own session.
DeleteTableOnFinish Set it to False if you do not want the created tables deleted after the load stops. This might be helpful when a large table is created upon the actor's first run, and then queries are made to that table.
UniformPartitionsCount The number of partitions created in test tables.
WorkloadType Type of load.
For Stock:
  • 0: InsertRandomOrder.
  • 1: SubmitRandomOrder.
  • 2: SubmitSameOrder.
  • 3: GetRandomCustomerHistory.
  • 4: GetCustomerHistory.
For Key-Value:
  • 0: UpsertRandom.
  • 1: InsertRandom.
  • 2: SelectRandom.
Workload Kind of load.
Stock:
  • ProductCount: Number of products.
  • Quantity: Quantity of each product in stock.
  • OrderCount: Initial number of orders in the database.
  • Limit: Minimum number of shards for tables.
Kv:
  • InitRowCount: Before load is generated, the load actor writes the specified number of rows to the table.
  • StringLen: Length of the value string.
  • ColumnsCnt: Number of columns to use in the table.
  • RowsCnt: Number of rows to insert or read per SQL query.

Examples

The following actor runs a stock load on the /slice/db database by making simple UPSERT queries of 64 threads during 30 seconds.

KqpLoad: {
    DurationSeconds: 30
    WindowDuration: 1
    WorkingDir: "/slice/db"
    NumOfSessions: 64
    UniformPartitionsCount: 1000
    DeleteTableOnFinish: 1
    WorkloadType: 0
    Stock: {
        ProductCount: 100
        Quantity: 1000
        OrderCount: 100
        Limit: 10
    }
}

As a result of the test, the number of successful transactions per second, the number of transaction execution retries, and the number of errors are output.

Previous
Overview
Next
KeyValueLoad
In this article: