Hologres:Stress testing data writes, updates, and point queries

Write and update mechanics

The tool supports two write modes:

• Fixed Copy mode -- Uses the PostgreSQL COPY statement, optimized with Fixed Plan for faster SQL execution.

• Insert mode -- Uses standard INSERT statements, also optimized with Fixed Plan.

In both modes, the tool automatically adds two columns to the table:

For every configured TEXT column, the tool writes a string of the specified length with the id value appended. The test stops when the target row count or duration is reached.

During updates, the same incrementing id pattern is used. For partial updates, only a subset of TEXT columns is rewritten.

Point query mechanics

The tool supports two query modes:

• Asynchronous mode -- The point query API is non-blocking. Multiple requests are batched into a single SQL statement, maximizing throughput. Best for high-throughput scenarios such as Flink-based real-time table joins.

• Synchronous mode -- The point query API is blocking. Each request maps to one SQL statement and completes before the next starts. Best for latency-sensitive scenarios.

During the test, the tool randomly selects a primary key within the configured range and performs a point query. The test stops when the target duration is reached.

About the test tool

holo-e2e-performance-tool is an open-source tool developed by the Hologres team. It integrates table creation, test data generation, and performance measurement into a single JAR, so there is no need to prepare test data separately. The primary key for generated data is a series of consecutive integers, which guarantees full hit rates during update and point query tests.

Step 1: Create the configuration file

On the ECS instance, create a file named test_insert.conf with the following content:

# Connection configuration
holoClient.jdbcUrl=jdbc:hologres://<ENDPOINT>:<PORT>/<DBNAME>
holoClient.username=<AccessKey_ID>
holoClient.password=<AccessKey_Secret>
holoClient.writeThreadSize=100

# Write configuration
put.threadSize=8
put.testByTime=false
put.rowNumber=200000000
put.testTime=600000

# Table configuration
put.tableName=kv_test
put.columnCount=20
put.columnSize=20
put.orientation=row

# Other configurations
put.createTableBeforeRun=true
put.deleteTableAfterDone=false
put.vacuumTableBeforeRun=false

This example creates a row-oriented table. To test a column-oriented or row-column hybrid table, change orientation to column or row,column.

Replace the placeholders with your actual values:

Configuration parameters

Step 2: Run the test

# Fixed Copy mode
java -jar holo-e2e-performance-tool-1.0.0.jar test_insert.conf FIXED_COPY

# Or, Insert mode
java -jar holo-e2e-performance-tool-1.0.0.jar test_insert.conf INSERT

Step 3: View the results

cat result.csv

See Result fields for the output format.

Global update

1. Create a file named test_update.conf. Use the same configuration as the write test, but change createTableBeforeRun to false:

   The only change from the write test configuration is createTableBeforeRun=false. This preserves the existing table and data.

      # Connection configuration
      holoClient.jdbcUrl=jdbc:hologres://<ENDPOINT>:<PORT>/<DBNAME>
      holoClient.username=<AccessKey_ID>
      holoClient.password=<AccessKey_Secret>
      holoClient.writeThreadSize=100
   
      # Write configuration
      put.threadSize=8
      put.testByTime=false
      put.rowNumber=200000000
      put.testTime=600000
   
      # Table configuration
      put.tableName=kv_test
      put.columnCount=20
      put.columnSize=20
      put.orientation=row
   
      # Other configurations
      put.createTableBeforeRun=false
      put.deleteTableAfterDone=false
      put.vacuumTableBeforeRun=false

2. Run the test:

      java -jar holo-e2e-performance-tool-1.0.0.jar test_update.conf FIXED_COPY

3. View the results:

      cat result.csv

Partial update

A partial update writes to a subset of the TEXT columns. The writeColumnCount parameter controls how many columns are updated.

1. Create a file named test_update_part.conf. Add the writeColumnCount parameter to the table configuration section:

   writeColumnCount=10 updates 10 of the 20 TEXT columns (50%). Adjust this ratio to match your workload.

      # Connection configuration
      holoClient.jdbcUrl=jdbc:hologres://<ENDPOINT>:<PORT>/<DBNAME>
      holoClient.username=<AccessKey_ID>
      holoClient.password=<AccessKey_Secret>
      holoClient.writeThreadSize=100
   
      # Write configuration
      put.threadSize=8
      put.testByTime=false
      put.rowNumber=200000000
      put.testTime=600000
   
      # Table configuration
      put.tableName=kv_test
      put.columnCount=20
      put.columnSize=20
      put.writeColumnCount=10
      put.orientation=row
   
      # Other configurations
      put.createTableBeforeRun=false
      put.deleteTableAfterDone=false
      put.vacuumTableBeforeRun=false

2. Run the test:

      java -jar holo-e2e-performance-tool-1.0.0.jar test_update_part.conf FIXED_COPY

3. View the results:

      cat result.csv

Step 1: Create the configuration file

Create a file named test_get.conf:

# Connection configuration
holoClient.jdbcUrl=jdbc:hologres://<ENDPOINT>:<PORT>/<DBNAME>
holoClient.username=<AccessKey_ID>
holoClient.password=<AccessKey_Secret>
holoClient.readThreadSize=32

# Test configuration
get.threadSize=8
get.testTime=300000
get.tableName=kv_test
get.async=true
get.vacuumTableBeforeRun=true
get.keyRangeParams=L1-200000000

# Table initialization configuration (used only in PREPARE_GET_DATA mode)
prepareGetData.rowNumber=200000000
prepareGetData.orientation=row
put.columnCount=20
put.columnSize=20

This example configures asynchronous point queries on a row-oriented table. To test synchronous queries, set get.async=false and adjust threadSize and readThreadSize accordingly (see the parameter table below).

Point query parameters

Step 2: Run the test

If the table already has data from the write or update tests, skip data preparation and run the point query directly:

# (Optional) Prepare data -- only if the table has no data
java -jar holo-e2e-performance-tool-1.0.0.jar test_get.conf PREPARE_GET_DATA

# Run the point query test
java -jar holo-e2e-performance-tool-1.0.0.jar test_get.conf GET

To use your own business data instead of generated data, the table must have a single-column primary key of the INT or BIGINT type with consecutive values.

Step 3: View the results

cat result.csv

Choose a storage format

Before reviewing the numbers, use this guide to select the right storage format for your workload:

20 columns

50 columns

100 columns

Key takeaways

• Row-oriented tables deliver the highest write throughput -- roughly 3x faster than column-oriented tables for the same column count.

• Partial updates outperform global updates on row-oriented tables because fewer columns are rewritten. On column-oriented and row-column hybrid tables, global updates may match or exceed partial update throughput.

• Point query latency is similar between row-oriented and row-column hybrid tables, typically under 10 ms for 20-column tables.

• Asynchronous point queries achieve higher QPS than synchronous queries, but at slightly higher latency.

• Write and update throughput decreases as column count increases -- a 100-column row-oriented table achieves roughly 28% of the write QPS of a 20-column table. Point query throughput is less affected, retaining 50-60% of QPS at 100 columns compared to 20 columns.

ECS client becomes the bottleneck

Symptom: QPS plateaus despite low CPU utilization on the Hologres instance.

Cause: The ECS instance running the test tool has reached its CPU or network bandwidth limit.

Solution: Monitor ECS CPU and network metrics during the test. If either is saturated, upgrade to a larger ECS instance type or reduce threadSize.

Lower-than-expected QPS

Symptom: QPS is significantly lower than the benchmark results listed above.

Possible causes and solutions:

1. Insufficient concurrency -- If Hologres CPU utilization is low, increase threadSize (for writes and updates) or readThreadSize (for point queries).

2. Instance has been upgraded or downgraded -- Performance characteristics differ between fresh and resized instances. Create a new instance for benchmarking.

3. Network latency -- Confirm that the ECS instance and the Hologres instance are in the same VPC and zone.

4. Data skew -- If using custom data, uneven distribution key values can concentrate load on a subset of shards.

Connection errors

Symptom: The test tool fails with connection-related exceptions.

Solution: Verify the JDBC connection string, AccessKey ID, and AccessKey secret. Confirm that the Hologres instance allows connections from the ECS instance's VPC.