Deadlock condition in driver

From: Daniel Migowski <dmigowski(at)ikoffice(dot)de>
To: pgsql-bugs(at)postgresql(dot)org
Subject: Deadlock condition in driver
Date: 2008-05-21 20:10:47
Message-ID: 483481C7.6030301@ikoffice.de
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-bugs

Hello dear developers,

I came across a deadlock condition in the JDBC driver that rises when
very large queries, containing thousends of statements are send to the
server with statement.execute(). I need this functionality, which work
well with less than 6445 statements to send update scripts along with my
application to the server.

The problem is the result of filled (and unemptied) TCP-Buffers. The
driver takes all statements and sends them to the server with the
extended query protocol. See this log snipped, which contains the last
10 lines before the eternal halt of the test application below (repeats
every 4 lines):

21:12:22.919 (1) FE=> Parse(stmt=null,query="select 0",oids={})
21:12:22.919 (1) FE=> Bind(stmt=null,portal=null)
21:12:22.919 (1) FE=> Describe(portal=null)
21:12:22.919 (1) FE=> Execute(portal=null,limit=0)
21:12:22.919 (1) FE=> Parse(stmt=null,query="select 0",oids={})
21:12:22.919 (1) FE=> Bind(stmt=null,portal=null)
21:12:22.919 (1) FE=> Describe(portal=null)
21:12:22.919 (1) FE=> Execute(portal=null,limit=0)
21:12:22.919 (1) FE=> Parse(stmt=null,query="select 0",oids={})
21:12:22.919 (1) FE=> Bind(stmt=null,portal=null)

Each statement sent to the server result in the following (yet still
unreceived) answers send from the server to the client:

21:27:50.169 (1) <=BE CommandStatus(SELECT)
21:27:50.169 (1) <=BE ParseComplete [null]
21:27:50.169 (1) <=BE BindComplete [null]
21:27:50.169 (1) <=BE RowDescription(1)
21:27:50.169 (1) <=BE DataRow

Since the driver is happy sending stuff, and the server happy answering
it, after a while the clients TCP receive buffer is full, some millis
later the servers TCP send buffer, some millies later the servers TCP
receive buffer and then finally the client TCP send buffer. Increasing
any of them delays the problem to a larger amount of statements.

When piping my script to psql, or sending it by PGAdmin there is no problem.

I suggest the following solution:

After sending 32kb (or the current send buffer size,
Socket.getSendBufferSize()), the client checks for the amount of data in
the receive buffer for every statement following. If its full (which
means the server might already be blocked), there are two possiblities,
from which the first is my favorite, and the second the option to
choose, if the first is not supported on the platform.

1. The receive buffer is increased in 32kb steps (or the current send
buffer size, or even 64k, taking the servers expected receive buffer
into account, to). This would unblock the server and gives enough space
for more (small) responses from the server. Afterwards the receive
buffer should be trimmed to its original size to don't become trapped by
decreased performance from large buffers. This method might be a bit
slower than the currently implemented one for cases in which the
answers of 32kb of statements would fit into the existing buffers, but
you don't expect lightning speed in such situations anyway, and
everything is better then being deadlocked in a production system.
2. We take all data available in the receive buffer already and place it
into a local byte[] to make space for more data, and release blocking on
the server side this way. This option might be a bit slower than the first.

A deadlock condition can yet be artificially constructed: If one sends a
SELECT that sends large amounts of data from the server to the client,
so the server is still not ready to process the next commands when we
continue to bombard it with statements, there might still be a lock
condition. But I cannot think of any use case where you have large
quantities of statements like me, AND expect to receive large amounts of
data which is discard anyway since there are following statements. So we
can ignore this case, and happily send gig sized update scripts to the
server in one call the driver :).

Currently my solution is to increase the TCP buffer size in a copy of
the PGStream class, which shadows the original class in the driver jar
(bad style...).

With best regards,
Daniel Migowski

------------- TEST APP --------------------

import java.io.IOException;
import java.sql.Connection;
import java.sql.DriverManager;
import java.sql.ResultSet;
import java.sql.SQLException;
import java.sql.Statement;
import java.util.Properties;

/**
* This class demonstrates the deadlock condition in the JDBC driver.
*
* Author: dmigowski
*/
public class DeadLockDemonstration {

public static void main (String args[]) throws IOException,
SQLException {
DriverManager.registerDriver(new org.postgresql.Driver());
Properties props = new Properties();
props.setProperty("user","ik");
props.setProperty("password","ik0000");
props.setProperty("loglevel","2");
Connection c = DriverManager.getConnection(
"jdbc:postgresql://localhost:5434/postgres",props);
StringBuilder b = new StringBuilder();
for(int i=1; i<100000; i++ ) {
b.append("select 0;");
}
Statement s = c.createStatement();
ResultSet r = s.executeQuery(b.toString());
}

}

Responses

Browse pgsql-bugs by date

  From Date Subject
Next Message Kris Jurka 2008-05-21 20:45:21 Re: Deadlock condition in driver
Previous Message Dave Page 2008-05-21 18:05:46 Re: Getting the command line to work