Skip to content

Latest commit

 

History

History
132 lines (110 loc) · 4.95 KB

pgbench.md

File metadata and controls

132 lines (110 loc) · 4.95 KB

PGAdapter - pgbench

pgbench can be used with PGAdapter, but with some limitations.

Follow these steps to initialize and run benchmarks with pgbench with PGAdapter:

Create Data Model

The default data model that is generated by pgbench does not include primary keys for the tables. Cloud Spanner requires all tables to have primary keys. Execute the following command to manually create the data model for pgbench:

psql -h /tmp -p 5432 -d my-database \
        -c "START BATCH DDL;
            CREATE TABLE pgbench_accounts (
                aid integer primary key  NOT NULL,            
                bid integer  NULL,                
                abalance integer  NULL,           
                filler varchar(84)  NULL
            );
            CREATE TABLE pgbench_branches (
                bid integer primary key  NOT NULL,            
                bbalance integer  NULL,           
                filler varchar(88)  NULL
            );
            CREATE TABLE pgbench_history (       
                tid integer  NOT NULL DEFAULT -1,                      
                bid integer  NOT NULL DEFAULT -1,                      
                aid integer  NOT NULL DEFAULT -1,                      
                delta integer  NULL,                    
                mtime timestamptz  NULL,
                filler varchar(22)  NULL,
                primary key (tid, bid, aid)
            );
            CREATE TABLE pgbench_tellers (
                tid integer  primary key NOT NULL,           
                bid integer  NULL,               
                tbalance integer  NULL,          
                filler varchar(84)  NULL
            );
            RUN BATCH;"

Initialize Data

pgbench deletes and inserts data into PostgreSQL using a combination of truncate, insert and copy statements. These statements all run in a single transaction. The amount of data that is modified during this transaction will exceed the transaction mutation limits of Cloud Spanner. This can be worked around by adding the following options to the pgbench initialization command:

pgbench "host=/tmp port=5432 dbname=my-database \
        options='-c spanner.force_autocommit=on -c spanner.autocommit_dml_mode=\'partitioned_non_atomic\''" \
        -i -Ig \
        --scale=100

These additional options do the following:

  1. spanner.force_autocommit=true: This instructs PGAdapter to ignore any transaction statements and execute all statements in autocommit mode. This prevents the initialization from being executed as a single, large transaction.
  2. spanner.autocommit_dml_mode='partitioned_non_atomic': This instructs PGAdapter to use Partitioned DML for (large) update statements. This ensures that a single statement succeeds even if it would exceed the transaction limits of Cloud Spanner, including large copy operations.
  3. -i activates initialization mode of pgbench.
  4. -Ig instructs pgbench to generate test data client side. PGAdapter does not support generating test data server side.

Running Benchmarks

You can run different benchmarks after finishing the steps above.

Default Benchmark

Run a default benchmark to verify that everything works as expected.

It is recommended to always use either --protocol=extended or --protocol=prepared for Cloud Spanner. The default --protocol=simple mode that is used by pgbench, will include all query parameters as literals in the query string and cause additional query parsing time on Cloud Spanner.

pgbench "host=/tmp port=5432 dbname=my-database" --protocol=extended

Number of Clients

Increase the number of clients and threads to increase the number of parallel transactions.

pgbench "host=/tmp port=5432 dbname=my-database" \
        --client=100 --jobs=100 \
        --progress=10 \
        --protocol=extended

Note that PGAdapter by default uses a gRPC channel pool with 4 channels. Running pgbench with more than 400 clients efficiently requires more channels.

Starting PGAdapter using Java with a larger gRPC channel pool:

java -jar pgadapter.jar -p my-project -i my-instance -d my-database \
     -r="numChannels=16"

Starting PGAdapter using Docker with a larger gRPC channel pool:

docker run -p 5432:5432 \
  gcr.io/cloud-spanner-pg-adapter/pgadapter \
  -p my-project -i my-instance -x \
  -r="numChannels=16"
pgbench "host=/tmp port=5432 dbname=my-database" \
        --client=800 --jobs=100 \
        --progress=10 \
        --protocol=extended

Dropping Tables

Execute the following command to remove the pgbench tables from your database if you no longer need them.

psql -h /tmp -p 5432 -d my-database \
        -c "START BATCH DDL;
            DROP TABLE pgbench_history;
            DROP TABLE pgbench_tellers;
            DROP TABLE pgbench_branches;
            DROP TABLE pgbench_accounts;
            RUN BATCH;"