YCSB/cloudspanner

Cloud Spanner Driver for YCSB

This driver provides a YCSB workload binding for Google's Cloud Spanner database, the first relational database service that is both strongly consistent and horizontally scalable. This binding is implemented using the official Java client library for Cloud Spanner which uses GRPC for making calls.

For best results, we strongly recommend running the benchmark from a Google Compute Engine (GCE) VM.

Running a Workload

We recommend reading the general guidelines in the YCSB documentation, and following the Cloud Spanner specific steps below.

1. Set up Cloud Spanner with the Expected Schema

Follow the Quickstart instructions in the Cloud Spanner documentation to set up a Cloud Spanner instance, and create a database with the following schema:

CREATE TABLE usertable (
  id STRING(MAX),
  field0 STRING(MAX),
  field1 STRING(MAX),
  field2 STRING(MAX),
  field3 STRING(MAX),
  field4 STRING(MAX),
  field5 STRING(MAX),
  field6 STRING(MAX),
  field7 STRING(MAX),
  field8 STRING(MAX),
  field9 STRING(MAX),
) PRIMARY KEY(id);

Make note of your project ID, instance ID, and database name.

2. Set Up Your Environment and Auth

Follow the set up instructions in the Cloud Spanner documentation to set up your environment and authentication. When not running on a GCE VM, make sure you run gcloud auth application-default login.

3. Edit Properties

In your YCSB root directory, edit cloudspanner/conf/cloudspanner.properties and specify your project ID, instance ID, and database name.

4. Run the YCSB Shell

Start the YCBS shell connected to Cloud Spanner using the following command:

./bin/ycsb shell cloudspanner -P cloudspanner/conf/cloudspanner.properties

You can use the insert, read, update, scan, and delete commands in the shell to experiment with your database and make sure the connection works. For example, try the following:

insert name field0=adam
read name field0
delete name

5. Load the Data

You can load, say, 10 GB of data into your YCSB database using the following command:

./bin/ycsb load cloudspanner -P cloudspanner/conf/cloudspanner.properties -P workloads/workloada -p recordcount=10000000 -p cloudspanner.batchinserts=1000 -threads 10 -s

We recommend batching insertions so as to reach ~1 MB of data per commit request; this is controlled via the cloudspanner.batchinserts parameter which we recommend setting to 1000 during data load.

If you wish to load a large database, you can run YCSB on multiple client VMs in parallel and use the insertstart and insertcount parameters to distribute the load as described here. In this case, we recommend the following:

• Use ordered inserts via specifying the YCSB parameter insertorder=ordered;
• Use zero-padding so that ordered inserts are actually lexicographically ordered; the option zeropadding = 12 is set in the default cloudspanner.properties file;
• Split the key range evenly between client VMs;
• Use few threads on each client VM, so that each individual commit request contains keys which are (close to) consecutive, and would thus likely address a single split; this also helps avoid overloading the servers.

The idea is that we have a number of 'write heads' which are all writing to different parts of the database (and thus talking to different servers), but each individual head is writing its own data (more or less) in order. See the best practices page for further details.

6. Run a Workload

After data load, you can a run a workload, say, workload B, using the following command:

./bin/ycsb run cloudspanner -P cloudspanner/conf/cloudspanner.properties -P workloads/workloadb -p recordcount=10000000 -p operationcount=1000000 -threads 10 -s 

Make sure that you use the same insertorder (i.e. ordered or hashed) and zeropadding as specified during the data load. Further details about running workloads are given in the YCSB wiki pages.

Configuration Options

In addition to the standard YCSB parameters, the following Cloud Spanner specific options can be configured using the -p parameter or in cloudspanner/conf/cloudspanner.properties.

• cloudspanner.database: (Required) The name of the database created in the instance, e.g. ycsb-database.
• cloudspanner.instance: (Required) The ID of the Cloud Spanner instance, e.g. ycsb-instance.
• cloudspanner.project: The ID of the project containing the Cloud Spanner instance, e.g. myproject. This is not strictly required and can often be automatically inferred from the environment.
• cloudspanner.readmode: Allows choosing between the read and query interface of Cloud Spanner. The default is query.
• cloudspanner.batchinserts: The number of inserts to batch into a single commit request. The default value is 1 which means no batching is done. Recommended value during data load is 1000.
• cloudspanner.boundedstaleness: Number of seconds we allow reads to be stale for. Set to 0 for strong reads (default). For performance gains, this should be set to 10 seconds.