pgworkload 0.1.12

Workload framework for the PostgreSQL protocol

pgworkload - workload utility for the PostgreSQL protocol

Overview

The goal of pgworkload is to ease the creation of workload scripts by providing a utility with the most common functionality already implemented.

pgworkload is run in conjunction with a user supplied Python class. This class defines the workload transactions and flow.

The user has complete control of what statements the transactions actually execute, and what transactions are executed in which order.

pgworkload can seed a database with random generated data, whose definition is supplied in a YAML file and can be extracted from a DDL SQL file.

Example

Class Bank in file workloads/bank.py is an example of one such user-created workload. The class defines 3 simple transactions that have to be executed by pgworkload. Have a look at the bank.py, bank.yaml and bank.sql in the workload folder in this project.

Head to file workload/bank.sql to see what the database schema look like. We have 2 tables:

• the transactions table, where we record the bank payment transactions.
• the ref_data table.

Take a close look at this last table: each column represent a different type, which brings us to the next file.

File bank.yaml is the data generation definition file. For each column of table ref_data, we deterministically generate random data. This file is meant as a guide to show what type of data can be generated, and what args are required.

File bank.py defines the workload. The workload is defined as a class object. The class defines 2 methods: run() and the constructor, __init__(). All other methods are part of the application logic of the workload. Read the comments along the code for more information.

Let's run the sample Bank workload.

Step 0 - env setup

# upgrade pip - must have pip version 20.3+ 
pip3 install --upgrade pip

pip3 install pgworkload

mkdir workloads
cd workloads

# the workload class
wget https://raw.githubusercontent.com/fabiog1901/pgworkload/main/workloads/bank.py

# the DDL file
wget https://raw.githubusercontent.com/fabiog1901/pgworkload/main/workloads/bank.sql

# the data generation definition file
wget https://raw.githubusercontent.com/fabiog1901/pgworkload/main/workloads/bank.yaml

Step 1 - init the workload

Make sure your CockroachDB cluster or PostgreSQL server is up and running.

Connect to the SQL prompt and load the bank.sql file. In CockroachDB, you can run

sql> \i bank.sql

Next, generate some CSV data to seed the database:

pgworkload util csv -i bank.yaml -x 1

The CSV files will be located inside a bank directory.

$ ls -lh bank
total 1032
-rw-r--r--  1 fabio  staff   513K Apr  9 13:01 ref_data.0_0_0.csv

Now you can import the CSV file. In CockroachDB, my favorite method is to use a webserver to serve the CSV file. Open a new terminal then start a simple python server

cd workloads
cd bank
python3 -m http.server 3000

If you open your browser at http://localhost:3000 you should see file ref_data.0_0_0.csv being served.

At the SQL prompt, import the file

sql> IMPORT INTO ref_data CSV DATA ('http://localhost:3000/ref_data.0_0_0.csv') WITH delimiter = e'\t'; 

Step 2 - Run the workload

Run the workload using 8 connections for 120 seconds or 100k cycles, whichever comes first.

# CockroachDB
pgworkload run -w bank.py -c 8 --url 'postgres://root@localhost:26257/bank?sslmode=disable&application_name=Bank' -d 120 -i 100000

# PostgreSQL
pgworkload run -w bank.py -c 8 --url 'postgres://root@localhost:5432/bank?sslmode=disable&application_name=Bank' -d 120 -i 100000

pgworkload uses exclusively the excellent Psycopg 3 to connect. No other ORMs or drivers/libraries are used. Psycopg has a very simple, neat way to create connections and execute statements and transactions.

pgworkload will output something like below

2022-01-28 17:22:43,893 [INFO] (MainProcess 29511) URL: 'postgres://root@localhost:26257/bank?sslmode=disable&application_name=Bank'
id               elapsed    tot_ops    tot_ops/s    period_ops    period_ops/s    mean(ms)    p50(ms)    p90(ms)    p95(ms)    p99(ms)    pMax(ms)
-------------  ---------  ---------  -----------  ------------  --------------  ----------  ---------  ---------  ---------  ---------  ----------
__cycle__             10       1342       133.72          1342           134.2       54.9       35.76     165.94     192.89     245.42      333.6
read                  10       1215       121.03          1215           121.5       41.11      19.58     113.21     146.79     208.86      291.02
txn1_new              10        130        12.95           130            13         48.29      53.81      74.7       90.84      95.66      108.37
txn2_verify           10        129        12.85           129            12.9       70.9       73.73      94.3       99.69     137.99      164.96
txn3_finalize         10        127        12.65           127            12.7       67.21      72.48      93.64     105.97     129.57      166 

[...]

2022-01-28 17:24:44,765 [INFO] (MainProcess 29511) Requested iteration/duration limit reached. Printing final stats
id               elapsed    tot_ops    tot_ops/s    period_ops    period_ops/s    mean(ms)    p50(ms)    p90(ms)    p95(ms)    p99(ms)    pMax(ms)
-------------  ---------  ---------  -----------  ------------  --------------  ----------  ---------  ---------  ---------  ---------  ----------
__cycle__            121      14519       120.12            66             6.6       94.08      96.68     203.74     216.83     242.24      262.69
read                 121      13050       107.96            54             5.4       70.6       62.7      127.88     151.29     203.52      203.62
txn1_new             121       1469        12.15             7             0.7       51.08      51.07      71.71      73.66      75.23       75.62
txn2_verify          121       1469        12.15            11             1.1       70.52      76.92     102.31     102.32     102.33      102.33
txn3_finalize        121       1469        12.15            12             1.2       81.19      98.97     103.88     103.97     103.98      103.98 

There are many built-in options. Check them out with

pgworkload --help

How it works

It’s helpful to understand first what pgworkload does:

• At runtime, pgworkload first imports the class you pass, bank.py.
• It spawns n threads for concurrent execution (see next section on Concurrency).
• By default, it sets the connection to autocommit mode.
• psycopg v3 will PREPARE statements automatically after 5 executions.
• Each thread creates a database connection - no need for a connection pool.
• In a loop, pgworkload will:
  • execute function run() which returns a list of functions.
  • execute each function in the list sequentially. Each function, typically, executes a SQL statement/transaction.
• Execution stats are funneled back to the MainThread, which aggregates and prints them to stdout.
• If the connection drops, it will recreate it. You can also program how long you want the connection to last.
• pgworkload stops once a limit has been reached (iteration/duration), or you Ctrl+C.

Concurrency - processes and threads

pgworkload uses both the multiprocessing and threading library to achieve high concurrency, that is, opening multiple connections to the DBMS.

There are 2 parameters that can be used to configure how many processes you want to create, and for each process, how many threads:

• --procs, or -x, to configure the count of processes (defaults to the CPU count)
• --concurrency, or -c, to configure the total number of executing workloads to run (also referred to as executing threads)

pgworkload will spread the load across the processes, so that each process has an even amount of threads.

Example: if we set --procs 4 and --concurrency 10, pgworkload will create as follows:

• Process-1: MainThread + 1 extra threads. Total = 2
• Process-2: MainThread + 1 extra threads. Total = 2
• Process-3: MainThread + 2 extra thread. Total = 3
• Process-4: MainThread + 2 extra thread. Total = 3

Total workloads = 10

This allows you to fine tune the count of Python processes and threads to fit your system.

Furthermore, each executing thread receives a unique ID (an integer). The ID is passed to the workload class with function setup(), along with the total count of threads, i.e. the value passed to -c/--concurrency. You can leverage the ID and the thread count in various ways, for example, to have each thread process a subset of a dataset.

Generating CSV files

• You can seed a database quickly by letting pgworkload generate pseudo-random data and import it.
• pgworkload takes the DDL as an input and creates an intermediate YAML file, with the definition of what data you want to create (a string, a number, a date, a bool..) based on the column data type.
• You then refine the YAML file to suit your needs, for example, the size of the string, a range for a date, the precision for a decimal, a choice among a discrete list of values..
• You can also specify what is the percentage of NULL for any column, or how many elements in an ARRAY type.
• You then specify the total row count, how many rows per file, and in what order, if any, to sort by.
• Then pgworkload will generate the data into CSV or TSV files, compress them if so requested.
• You can then optionally merge-sort the files using command merge.

Write up blog: Generate multiple large sorted csv files with pseudo-random data

Find out more on the yaml, csv and merge commands by running

pgworkload util --help

Consult file workloads/bank.yaml for a list of all available generators and options.

Built-in Workloads

pgworkload has the following workload already built-in and can be called without the need to pass a class file

Querybench

Querybench runs a list of SQL Statements sequentially and iteratively. It assumes the schema and data have been created and loaded.

SQL statements file mystmts.sql

-- Query 1
select 1;
select 
  version();
-- select now();

-- Query 2
SELECT * FROM my_table 
WHERE id = 1234;

Run Querybench like this:

pgworkload run --builtin-workload Querybench --args mystmts.sql --url <conn-string>

Acknowledgments

Some methods and classes have been taken and modified from, or inspired by, https://github.com/cockroachdb/movr