client socket TIME_WAIT state after PQfinish

2010/9/27 Guillaume Du Pasquier <guillaume.dupasquier@sensometrix.ch>:

In both cases, the client socket (pgadmin or my program) remains in
TIME_WAIT state.

I have used wireshark to sniff the TCP protocol.

We have at the end of a connection:

Client                  Server

   --->    FIN,ACK  --->
   <---    FIN,ACK  <---
   --->      ACK    --->

This ends up in a TIME_WAIT state. The TCP protocol should be

According to the Two Generals' Problem [1]<URL:http://en.wikipedia.org/wiki/Two_Generals&#39;_Problem&gt;, one of the sides
necessarily doesn't know whether the other side has received its last
packet. Therefore, TCP lets one of the sides sit in TIME_WAIT status
for as long as any packets could in principle survive on the network
(typically defined as 2 minutes on IP networks IIRC), and potentially
disturb a new connection between the same (dst IP, dst port, src IP,
src port) combination.

[1]: <URL:http://en.wikipedia.org/wiki/Two_Generals&#39;_Problem&gt;

Client                  Server

   --->    FIN,ACK  --->
   <---      ACK    <---
   <---    FIN,ACK  <---

AFAIK, this last ACK (in above packet) is not needed: the server can
ACK the client's FIN _while_ it sends its own FIN (by using an
appropriate sequence number, as FIN "uses" one byte in the sequence).

   --->      ACK    --->

I suppose there is a bug in the postgresql server that do not send an ack to
the client.

I don't think so.

Nicolas

On Mon, Sep 27, 2010 at 12:22 PM, Nicolas Barbier
<nicolas.barbier@gmail.com> wrote:

According to the Two Generals' Problem [1], one of the sides
necessarily doesn't know whether the other side has received its last
packet. Therefore, TCP lets one of the sides sit in TIME_WAIT status
for as long as any packets could in principle survive on the network
(typically defined as 2 minutes on IP networks IIRC), and potentially
disturb a new connection between the same (dst IP, dst port, src IP,
src port) combination.

In other words, this is the way TCP is designed to work, not something
specific to PostgreSQL.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise Postgres Company

Dear Nicolas, Dear Robert,

Thank you for your quick answers.
We do not have such behavior using SSL, how do you explain it ?
I suppose that openssl is using the setsockopt SO_LINGER that
removes this behavior. Therefore, there is a RST sent to close
the socket.

We work with an environment that uses a lot of socket connections.
Therefore, many file descriptors are opened. If after each
Sql requests a TIME_WAIT arises we will end up with many
file descriptors opened. By default the maximum number of file descriptors
is set to 1024 and we reach that number quite fast.

Do you have any advices to get rid of this TIME_WAIT problem ?
Our client runs on the same machine as the postgresql server.
Would it be possible to use PF_UNIX sockets ?

Thank you,

Franck

-----Original Message-----
From: Nicolas Barbier [mailto:nicolas.barbier@gmail.com]
Sent: lundi 27 septembre 2010 18:23
To: Guillaume Du Pasquier
Cc: pgsql-hackers@postgresql.org
Subject: Re: [HACKERS] client socket TIME_WAIT state after PQfinish

2010/9/27 Guillaume Du Pasquier <guillaume.dupasquier@sensometrix.ch>:

In both cases, the client socket (pgadmin or my program) remains in
TIME_WAIT state.

I have used wireshark to sniff the TCP protocol.

We have at the end of a connection:

Client                  Server

   --->    FIN,ACK  --->
   <---    FIN,ACK  <---
   --->      ACK    --->

This ends up in a TIME_WAIT state. The TCP protocol should be

According to the Two Generals' Problem [1]<URL:http://en.wikipedia.org/wiki/Two_Generals&#39;_Problem&gt;, one of the sides
necessarily doesn't know whether the other side has received its last
packet. Therefore, TCP lets one of the sides sit in TIME_WAIT status
for as long as any packets could in principle survive on the network
(typically defined as 2 minutes on IP networks IIRC), and potentially
disturb a new connection between the same (dst IP, dst port, src IP,
src port) combination.

[1]: <URL:http://en.wikipedia.org/wiki/Two_Generals&#39;_Problem&gt;

Client                  Server

   --->    FIN,ACK  --->
   <---      ACK    <---
   <---    FIN,ACK  <---

AFAIK, this last ACK (in above packet) is not needed: the server can
ACK the client's FIN _while_ it sends its own FIN (by using an
appropriate sequence number, as FIN "uses" one byte in the sequence).

   --->      ACK    --->

I suppose there is a bug in the postgresql server that do not send an ack to
the client.

I don't think so.

Nicolas

On Mon, Sep 27, 2010 at 12:56 PM, Guillaume Du Pasquier
<guillaume.dupasquier@sensometrix.ch> wrote:

Dear Nicolas, Dear Robert,

Thank you for your quick answers.
We do not have such behavior using SSL, how do you explain it ?
I suppose that openssl is using the setsockopt SO_LINGER that
removes this behavior. Therefore, there is a RST sent to close
the socket.

We work with an environment that uses a lot of socket connections.
Therefore, many file descriptors are opened. If after each
Sql requests a TIME_WAIT arises we will end up with many
file descriptors opened. By default the maximum number of file descriptors
is set to 1024 and we reach that number quite fast.

Hrm. Does a socket in the TIME_WAIT state count against the number of
open file descriptors? Certainly, it shouldn't count against the
per-process limit, as the process has already closed it.

Do you have any advices to get rid of this TIME_WAIT problem ?
Our client runs on the same machine as the postgresql server.
Would it be possible to use PF_UNIX sockets ?

Yeah, actually that's the default, if you just run "psql" with no
parameters. It looks for a socket in /tmp/.s.PGSQL.5432.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise Postgres Company

2010/9/27 Robert Haas <robertmhaas@gmail.com>:

On Mon, Sep 27, 2010 at 12:56 PM, Guillaume Du Pasquier
<guillaume.dupasquier@sensometrix.ch> wrote:

Our client runs on the same machine as the postgresql server.
Would it be possible to use PF_UNIX sockets ?

Yeah, actually that's the default, if you just run "psql" with no
parameters.  It looks for a socket in /tmp/.s.PGSQL.5432.

It depends on the interface; e.g., the JDBC driver doesn't support
unix sockets, AFAIR.

Nicolas