GCSB
It's like YCSB but with more Google. A simple tool meant to generate load against Google Spanner databases. The primary goals of the project are
- Write randomized data to tables in order to
- Facilitate load testing
- Initiate database splits
- Generate read/write load against user provided schemas
Quickstart
To initiate a simple load test against your spanner instance using one of our test schemas
Create a test table
You can use your own schema if you'd prefer, but we provide a few test schemas to help you get started. To get started, create a table named SingleSingers
gcloud spanner databases ddl update YOUR_DATABASE_ID --instance=YOUR_INSTANCE_ID --ddl-file=schemas/single_table.sql
Load data into table
Load some data into the table to seed the upcoming laod test. In the example below, we are loading 10,000 rows of random data into the table SingleSingers
gcsb load -p YOUR_GCP_PROJECT_ID -i YOUR_INSTANCE_ID -d YOUR_DATABASE_ID -t SingleSingers -o 10000
Run a load test
Now you can perform a load test operation by using the run sub command. The below command will generate 10,000 operations. Of these operations, 75% will be READ operations and 25% will be writes. These operations will be performed over 50 threads.
gcsb run -p YOUR_GCP_PROJECT_ID -i YOUR_INSTANCE_ID -d YOUR_DATABASE_ID -t SingleSingers -o 10000 --reads 75 --writes 25 --threads 50
Operations
The tool usage is generally broken down into two categories, load and run operations.
Load
GCSB proides batched loading functionality to facilitate load testing as well as assist with performing database splits on your tables.
At runtime, GCSB will detect the schema of the table you're loading data for an create data generators that are appropriate for the column types you have in your database. Each type of generator has some configurable funcationality that allows you to refine the type, length, or range of data the tool generates. For in depth information on the various configuration values, please read the comments in example_gcsb.yaml.
Single table load
By default, GCSB will detect the table schema and create default random data generators based on the columns it finds. In order to tune the values the generator creates, you must create override configurations in the gcsb.yaml file. Please see that file's documentation for more information.
gcsb load -t TABLE_NAME -o NUM_ROWS
Additionally, please see gcsb load --help for additional configuration options.
Multiple table load
Similar to the above Single table load, you may specify multiple tables by repeating the -t TABLE_NAME argument. By default, the number of operations is applied to each table. For example, specifying 2 tables with 1000 operations, will yield 2000 total operations. 1000 per table.
gcsb load -t TABLE1 -t TABLE2 -o NUM_ROWS
Operations per table can be configured in the yaml configuration. For example
tables:
- name: TABLE1
operations:
total: 500
- name: TABLE2
operations:
total: 500
Loading into interleaved tables
Loading data into interleaved tables is supported but has some behavioral side effects that should be known. When loading data into an interleaved table, GCSB will detect all tables in the hierarchy and begin loading data at the familial apex. The configured number of operations applies to this apex table. By default, the number of tables for each child table is multiplied by 5. For example:
Using our test INTERLEAVE schema, we see an INTERLEAVE relationship between the Singers, Albums, and Songs tables.
If we execute a load operation against these tables with total operations set to 10 we will see the following occur
gcsb load -t Songs -o 10
+---------+------------+------+-------+---------+
| TABLE | OPERATIONS | READ | WRITE | CONTEXT |
+---------+------------+------+-------+---------+
| Singers | 10 | N/A | N/A | LOAD |
| Albums | 50 | N/A | N/A | LOAD |
| Songs | 250 | N/A | N/A | LOAD |
+---------+------------+------+-------+---------+
In this case, for each child table we take the number of operations for the parent and multiply it by the default value of 5.
To change this multiplier, we use the yaml configuration file for the table we want. The operations.total value becomes a multiplier.
tables:
- name: Albums
operations:
total: 10
- name: Songs
operations:
total: 20
At present, GCSB will sort it's operations from the apex down. Meaning it will populate the Singers table first and then it's child, and then the next child. Multiple table operations are not mixed within the same transaction.
Run
Single table run
By default, GCSB will detect the table schema and create default random data generators based on the columns it finds. In order to tune the values the generator creates, you must create override configurations in the gcsb.yaml file. Please see that file's documentation for more information.
gcsb run -p YOUR_GCP_PROJECT_ID -i YOUR_INSTANCE_ID -d YOUR_DATABASE_ID -t SingleSingers -o 10000 --reads 75 --writes 25 --threads 50
Additionally, please see gcsb run --help for additional configuration options.
Multiple table run
Similar to the above Single table run, you may specify multiple tables by repeating the -t TABLE_NAME argument. By default, the number of operations is applied to each table. For example, specifying 2 tables with 1000 operations, will yield 2000 total operations. 1000 per table.
gcsb run -t TABLE1 -t TABLE2 -o NUM_ROWS
Operations per table can be configured in the yaml configuration. For example
tables:
- name: TABLE1
operations:
total: 500
- name: TABLE2
operations:
total: 500
Running against interleaved tables
Run operations against interleaved tables are only supported on the APEX table.
Using our test INTERLEAVE schema, we see an INTERLEAVE relationship between the Singers, Albums, and Songs tables.
You will notice that if we try to run against any child table an error will occur.
gcsb run -t Songs -o 10
unable to execute run operation: can only execute run against apex table (try 'Singers')
Distributed testing
GCSB is intended to run in a stateless mannger. This design choice was to allow massive horizontal scaling of gcsb to stress your database to it's absolute limits. During development we've identified kubernetes as the prefered tool for the job. We've provided two separate tutorials for running gcsb inside of kubernetes
- GKE - For running GCSB inside GKE using a service account key. This can be used for non-GKE clusters as well as it contains instructions for mounting a service account key into the container.
- GKE with Workload Identity - For running GCSB inside a GKE cluster that has workload identity turned on. This is most useful in organizations that have security policies preventing you from generating or downloading a service account key.
Configuration
The tool can receive configuration input in several different ways. The tool will load the file gcsb.yaml if it detects it in the current working directory. Alternatively you can use the global flag -c to specify a path to the configuration file. Each sub-command has a number of configuration flags that are relevant to that operation. These values are bound to their counterparts in the yaml configuration file and take precedent over the config file. Think of them as overrides. The same is true for environment variables.
Please note, at present, the yaml conifguration file is the only way to specify generator overrides for data loading and write operations. Without this file, the tool will use a random data generator that is appropriate for the table schema it detects at runtime.
For in depth information on the various configuration values, please read the comments in example_gcsb.yaml
Roadmap
Not Supported (yet)
- Generating read operations utilizing ReadByIndex
- Generating NULL values for load operations. If a column is NULLable, gcsb will still generate a value for it.
- JSON column types
- STRUCT Objects.
- VIEWS
- Inserting data across multiple tables in the same transaction
- No SCAN or DELETE operations are supported at this time
- Tables with foreign key relationships
- Testing multiple tables at once
Development
Build
make build
Test
make test
