Thread: Dropped connections with pg_basebackup

Have an existing setup of 9.3 servers. Replication has been rock solid,
but recently the circuits between data centers were upgraded and
pg_basebackup now seems to fail often when setting up streaming
replication. What used to take 10+ hours now  only took 68 minutes, but
had to do many retries. Many attempts fail within minutes while others
go to 90% or higher and then drop. The reason we are doing a sync is
because we have to swap data centers every so often for compliance. So I
had to swap master and slave.

Calling pg_basebackup like this:
pg_basebackup -P -R -X s -h <HostName> -D <Folder> -U replicator

The error we keep having is:
Sep 23 13:36:32 <HostName> postgres[16804]: [11-1] 2015-09-23 13:36:32
EDT <IP> [unknown] replicator LOG: SSL error: bad write retry
Sep 23 13:36:32 <HostName> postgres[16804]: [12-1] 2015-09-23 13:36:32
EDT <IP> [unknown] replicator LOG: SSL error: bad write retry
Sep 23 13:36:32 <HostName> postgres[16804]: [13-1] 2015-09-23 13:36:32
EDT <IP> [unknown] replicator FATAL: connection to client lost
Sep 23 13:36:32 <HostName> postgres[16972]: [9-1] 2015-09-23 13:36:32
EDT <IP> [unknown] replicator LOG: could not receive data from client:
Connection reset by peer

I have been working with the network team and we have even been actively
monitoring the line, and running ping, as the replication is setup. At
the point the connection reset by peer error happens, we don't see any
issue with the network and ping doesn't show an issue at that point in time.

The issue also happened on another set of machines and likewise, had to
retry many times before pg_basebackup would do the initial sync. Once
the initial sync is set, replication is fine.

I  tried both "-X s" (stream) and "-X f" (fetch) and both fail often.

Any ideas what may be going on?

On 09/24/2015 12:57 PM, Francisco Reyes wrote:
> Have an existing setup of 9.3 servers. Replication has been rock solid,
> but recently the circuits between data centers were upgraded and
> pg_basebackup now seems to fail often when setting up streaming
> replication. What used to take 10+ hours now  only took 68 minutes, but
> had to do many retries. Many attempts fail within minutes while others
> go to 90% or higher and then drop. The reason we are doing a sync is
> because we have to swap data centers every so often for compliance. So I
> had to swap master and slave.
>
> Calling pg_basebackup like this:
> pg_basebackup -P -R -X s -h <HostName> -D <Folder> -U replicator
>
> The error we keep having is:
> Sep 23 13:36:32 <HostName> postgres[16804]: [11-1] 2015-09-23 13:36:32
> EDT <IP> [unknown] replicator LOG: SSL error: bad write retry
> Sep 23 13:36:32 <HostName> postgres[16804]: [12-1] 2015-09-23 13:36:32
> EDT <IP> [unknown] replicator LOG: SSL error: bad write retry

Seems to be an SSL problem, so how is your SSL set up on the servers?

> Sep 23 13:36:32 <HostName> postgres[16804]: [13-1] 2015-09-23 13:36:32
> EDT <IP> [unknown] replicator FATAL: connection to client lost
> Sep 23 13:36:32 <HostName> postgres[16972]: [9-1] 2015-09-23 13:36:32
> EDT <IP> [unknown] replicator LOG: could not receive data from client:
> Connection reset by peer
>
> I have been working with the network team and we have even been actively
> monitoring the line, and running ping, as the replication is setup. At
> the point the connection reset by peer error happens, we don't see any
> issue with the network and ping doesn't show an issue at that point in
> time.
>
> The issue also happened on another set of machines and likewise, had to
> retry many times before pg_basebackup would do the initial sync. Once
> the initial sync is set, replication is fine.
>
> I  tried both "-X s" (stream) and "-X f" (fetch) and both fail often.
>
> Any ideas what may be going on?
>
>


--
Adrian Klaver
adrian.klaver@aklaver.com

Sherrylyn Branchaw wrote:
> I'm assuming based on the "SSL error" that you have ssl set to 'on'. What's
> your ssl_renegotiation_limit? The default is 512MB, but setting it to 0 has
> solved problems for a number of people on this list, including myself.

Moreover, the default has been set to 0, because the bugs both in our
usage and in OpenSSL code itself seem never to end.  Just disable it.

--
Álvaro Herrera                http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

On 09/24/2015 04:29 PM, Sherrylyn Branchaw wrote:
> I'm assuming based on the "SSL error" that you have ssl set to 'on'.
> What's your ssl_renegotiation_limit? The default is 512MB, but setting
> it to 0 has solved problems for a number of people on this list,
> including myself.

I have also seen instances were ssl_renegotiation_limit=0 helped and I
already tried that. Did not help in this case.

Perhaps will try some tests with a non SSL connection. These are
machines in an internal network so it may not be too much a security
issue to turn off SSL at least during initial sync.

On 09/24/2015 04:34 PM, Alvaro Herrera wrote:
> Sherrylyn Branchaw wrote:
> Moreover, the default has been set to 0, because the bugs both in our
> usage and in OpenSSL code itself seem never to end.  Just disable it.


Set it to 0 and did not help.
Likely will move all machines to have it =0 since I have seen some SSL
errors in logs.