pgreplay - record and replay real-life database workloads

pgreplay reads a PostgreSQL log file (not a WAL file), extracts the SQL statements and executes them in the same order and with the original timing against a PostgreSQL database.

If the execution of statements gets behind schedule, warning messages are issued that indicate that the server cannot handle the load in a timely fashion.

A final report gives you a useful statistical analysis of your workload and its execution.

The idea is to replay a real-world database workload as exactly as possible.

This is useful for performance tests, particularly in the following situations:

• You want to compare the performance of your PostgreSQL application on different hardware or different operating systems.
• You want to upgrade your database and want to make sure that the new database version does not suffer from performance regressions that affect you.

Moreover, pgreplay can give you some feeling as to how your application might scale by allowing you to try to replay the workload at a higher speed (if that is possible; see implementation details below). Be warned, though, that 500 users working at double speed is not really the same as 1000 users working at normal speed.

While pgreplay will find out if your database application will encounter performance problems, it does not provide a lot of help in the analysis of the cause of these problems. Combine pgreplay with a specialized analysis program like pgBadger for that.

As an additional feature, pgreplay lets you split the replay in two parts: you can parse the log file and create a "replay file", which contains just the statements to be replayed and is hopefully much smaller than the original log file.
Such a replay file can then be run against a database.

pgreplay is written by Laurenz Albe and is inspired by "Playr" which never made it out of Beta.

Installation

pgreplay needs PostgreSQL 8.0 or better.

It is supposed to compile without warnings and run on all platforms supported by PostgreSQL.
Since I only got to test it on Linux, AIX, FreeBSD and Windows, there may be problems with other platforms. I am interested in reports and fixes for these platforms.
On Windows, only the MinGW build environment is supported (I have no other compiler). That means that there is currently no 64-bit build for Windows (but a 32-bit executable should work fine anywhere).

To build pgreplay, you will need the pg_config utility. If you installed PostgreSQL using installation packages, you will probably have to install the development package that contains pg_config and the header files.

If pg_config is on the PATH, the installation process will look like this:

• unpack the tarball
• ./configure
• make
• make test (optional, described below)
• make install (as superuser)

If your PostgreSQL installation is in a nonstandard directory, you will have to use the --with-postgres=<path to location of pg_config> option of configure.

Unless you link it statically, pgreplay requires the PostgreSQL client shared library on the system where it is run.

The following utilities are only necessary if you intend to develop pgreplay:

• autoconf 2.62 or better to generate configure
• GNU tar to make tarball (unless you want to roll it by hand)
• groff to make the HTML documentation with make html

Testing

You can run a test on pgreplay before installing by running make test. This will parse sample log files and check that the result is as expected.

Then an attempt is made to replay the log files and check if that works as expected. For this you need psql installed and a PostgreSQL server running (on this or another machine) so that the following command will succeed:

psql -U postgres -d postgres -l

You can set up the PGPORT and PGHOST environment variables and a password file for the user if necessary.

Usage

First, you will need to record your real-life workload. For that, set the following parameters in postgresql.conf:

• log_min_messages = error (or more)
  (if you know that you have no cancel requests, log will do)
• log_min_error_statement = log (or more)
• log_connections = on
• log_disconnections = on
• log_line_prefix = '%m|%u|%d|%c|' (if you don't use CSV logging)
• log_statement = 'all'
• lc_messages must be set to English (the encoding does not matter)
• bytea_output = escape (from version 9.0 on, only if you want to replay the log on 8.4 or earlier)

It is highly recommended that you use CSV logging, because anything that the PostgreSQL server or any loaded modules write to standard error will be written to the stderr log and might confuse the parser.

Then let your users have their way with the database.

Make sure that you have a pg_dumpall of the database cluster from the time of the start of your log file (or use the -b option with the time of your backup). Alternatively, you can use point-in-time-recovery to clone your database at the appropriate time.

When you are done, restore the database (in the "before" state) to the machine where you want to perform the load test and run pgreplay against that database.

Try to create a scenario as similar to your production system as possible (except for the change you want to test, of course). For example, if your clients connect over the network, run pgreplay on a different machine from where the database server is running.

Since passwords are not logged (and pgreplay consequently has no way of knowing them), you have two options: either change pg_hba.conf on the test database to allow trust authentication or (if that is unacceptable) create a password file as described by the PostgreSQL documentation. Alternatively, you can change the passwords of all application users to one single password that you supply to pgreplay with the -W option.

Limitations

pgreplay can only replay what is logged by PostgreSQL. This leads to some limitations:

• COPY statements will not be replayed, because the copy data are not logged. I could have supported COPY TO statements, but that would have imposed a requirement that the directory structure on the replay system must be identical to the original machine. And if your application runs on the same machine as your database and they interact on the file system, pgreplay will probably not help you much anyway.
• Fast-path API function calls are not logged and will not be replayed. Unfortunately, this includes the Large Object API.
• Since the log file is always written in the database encoding (which you can specify with the -E switch of pgreplay), all SET client_encoding statements will be ignored.
• If your cluster contains databases with different encoding, the log file will have mixed encoding as well. You cannot use pgreplay well in such an environment, because many statements against databases whose encoding does not match the -E switch will fail.
• Since the preparation time of prepared statements is not logged (unless log_min_messages is debug2 or more), these statements will be prepared immediately before they are first executed during replay.
• All parameters of prepared statements are logged as strings, no matter what type was originally specified during bind. This can cause errors during replay with expressions like $1 + $2, which will cause the error operator is not unique: unknown + unknown.

While pgreplay makes sure that commands are sent to the server in the order in which they were originally executed, there is no way to guarantee that they will be executed in the same order during replay: Network delay, processor contention and other factors may cause a later command to "overtake" an earlier one. While this does not matter if the commands don't affect each other, it can lead to SQL statements hitting locks unexpectedly, causing replay to deadlock and "hang". This is particularly likely if many different sessions change the same data repeatedly in short intervals.

You can work around this problem by canceling the waiting statement with pg_cancel_backend. Replay should continue normally after that.

Implementation details

pgreplay will track the "session ID" associated with each log entry (the session ID uniquely identifies a database connection). For each new session ID, a new database connection will be opened during replay. Each statement will be sent on the corresponding connection, so transactions are preserved and concurrent sessions cannot get in each other's way.

The order of statements in the log file is strictly preserved, so there cannot be any race conditions caused by different execution speeds on separate connections. On the other hand, that means that long running queries on one connection may stall execution on concurrent connections, but that's all you can get if you want to reproduce the exact same workload on a system that behaves differently.

As an example, consider this (simplified) log file:

session 1|connect
session 2|connect
session 1|statement: BEGIN
session 1|statement: SELECT something(1)
session 2|statement: BEGIN
session 2|statement: SELECT something(2)
session 1|statement: SELECT something(3)
session 2|statement: ROLLBACK
session 2|disconnect
session 1|statement: COMMIT
session 2|disconnect

This will cause two database connections to be opened, so the ROLLBACK in session 2 will not affect session 1. If SELECT something(2) takes longer than expected (longer than it did in the original), that will not stall the execution of SELECT something(3) because it runs on a different connection. The ROLLBACK, however, has to wait for the completion of the long statement. Since the order of statements is preserved, the COMMIT on session 1 will have to wait until the ROLLBACK on session 2 has started (but it does not have to wait for the completion of the ROLLBACK).

pgreplay is implemented in C and makes heavy use of asynchronous command processing (which is the reason why it is implemented in C). This way a single process can handle many concurrent connections, which makes it possible to get away without multithreading or multiprocessing.

This avoids the need for synchronization and many portability problems. But since TINSTAAFL, the choice of C brings along its own portability problems. Go figure.

Replay file format

The replay file is a binary file, integer numbers are stored in network byte order.

Each record in the replay file corresponds to one database operation and is constructed as follows:

• 4-byte unsigned int: log file timestamp in seconds since 2000-01-01
• 4-byte unsigned int: fractional part of log file timestamp in microseconds
• 8-byte unsigned int: session id
• 1-byte unsigned int: type of the database action:
  • 0 is connect
  • 1 is disconnect
  • 2 is simple statement execution
  • 3 is statement preparation
  • 4 is execution of a prepared statement
  • 5 is cancel request
• The remainder of the record is specific to the action, strings are stored with a preceeding 4-byte unsigned int that contains the length. Read the source for details.
• Each record is terminated by a new-line character (byte 0x0A).

Support

If you have a problem or question, the preferred option is to open an issue. This requires a GitHub account.

Professional support can be bought from Cybertec.

TODO list

Nothing currently. Tell me if you have good ideas.