[syslog-ng] syslog-ng 3.2.4 MySQL connection loss Server has gone away

[Balazs Scheidler]

Hi,

On Thu, 2011-09-29 at 22:02 -0700, Erik Maciejewski wrote:
> On Sep 29, 2011, at 1:06 AM, Balazs Scheidler wrote:
> 
> > On Thu, 2011-09-29 at 00:21 -0700, Erik Maciejewski wrote:
> >> On Sep 28, 2011, at 12:06 AM, Balazs Scheidler wrote:
> >> 
> >>> Hi,
> >>> 
> >>> Sorry for the late answer, but better late than never :)
> >>> 
> >>> On Sun, 2011-09-04 at 23:15 -0700, Erik Maciejewski wrote:
> >>>> Hello,
> >>>> 
> >>>> I have been researching an issue with syslog-ng 3.2.4 (both the Linux glibc2.3.6 and platform 
> >>>> independent compiled from source versions) and MySQL on CentOS 5.6 in which syslog-ng tries to 
> >>>> insert a log messages using a dead TCP or unix socket connection to a MySQL database. The issue 
> >>>> results in the max attempt to insert a log message and then the message subsequently dropped. 
> >>>> This seems to occur regularly in predominately low message volume environments, but has the 
> >>>> potential to affect all environments using a MySQL (or other db) destination. The root cause 
> >>>> of the issue seems to be directly related to the health of the connection between syslog-ng 
> >>>> and MySQL and can be affected by the "wait_timeout" value used by MySQL to kill off inactive 
> >>>> connections. I want to provide justification for implementing a fix to syslog-ng as I believe 
> >>>> manipulating a, possibly tuned, environment variable for MySQL would seem like the wrong 
> >>>> approach to take for remedial action.
> >>>> 
> >>>> I feel that many MySQL instances are implemented in shared environments and are themselves 
> >>>> shared by many applications. New connections to a MySQL instance are generally regarded as 
> >>>> low cost and, it would seem, more often than not never used in a persisted fashion when 
> >>>> supporting distributed applications. That being said, many times the "wait_timeout" 
> >>>> value is set very low so that the MySQL instance can effectively control the number of
> >>>> simultaneous connections.
> >>>> 
> >>>> 
> >>>> By taking somewhat of a black-box approach to the issue (I'm just starting to explore 
> >>>> syslog-ng), I noticed that there doesn't seem to be any health checking of the TCP or 
> >>>> unix socket connection in the SQL destination implementation after the initial 
> >>>> connection is made (most of the db actions being abstracted away by libdbi). Even 
> >>>> if a syslog-ng database thread is suspended and reactivated due to error, there 
> >>>> does not seem to be any attempt to check the health of the database connection.
> >>>> 
> >>>> I have implemented a fix in my compiled version that will check a database 
> >>>> connection's health and attempt to reconnect upon discovering a dead 
> >>>> connection. I would like to know the policies/procedures/best practices 
> >>>> for submitting such issues and/or fixes for discussion, review, and 
> >>>> implementation. If this is a known issue or I am just wrong, all the better! 
> >>> 
> >>>> Thank you guys for keeping this project going and I'm looking forward to 
> >>>> implementing syslog-ng in a production environment in the very near future!
> >>> 
> >>> 
> >>> syslog-ng doesn't explicitly check the aliveness of an SQL connection,
> >>> however it does drop and reestablish connection if an error occurs.
> >>> 
> >>> If you look at afsql_dd_database_thread() function in afsql.c, you'll
> >>> see this code:
> >>> 
> >>>     if (!afsql_dd_insert_db(self))
> >>>       {
> >>>         afsql_dd_disconnect(self);
> >>>         afsql_dd_suspend(self);
> >>>       }
> >>> 
> >>> afsql_dd_insert_db() should return FALSE for any failures, and syslog-ng
> >>> basically assumes that if the database side initiates the closure of a
> >>> connection, that'll trickle up to syslog-ng as an error return to
> >>> dbi_conn_query(), which will in turn cause the SQL destination to
> >>> suspend its operations for time_reopen() amount of time and then
> >>> reconnect.
> >>> 
> >>> What kind of fix did you implement yourself? Can you post the patch?
> >>> Thanks.
> >>> 
> >>> -- 
> >>> Bazsi
> >>> 
> >>> 
> >> No worries :)
> >> 
> >> Basically, I would propose explicitly checking the health of the connection 
> >> (using libdbi). By doing so, I believe there is a likelihood that log messages 
> >> will not be lost in the event that the connection to the database is terminated 
> >> (ex: db process restart or db connection TTL).
> >> 
> >> Towards the beginning of the function afsql_dd_insert_db() in afsql.c I have:
> >> 
> >>    if (!self->dbi_ctx)
> >>      {
> >> 	    /*handle new db connection*/
> >> 	
> >> From my understanding, this section of code checks to see if a valid connection
> >> instance exists. If not, a new instance is created. I'm not too familiar with 
> >> libdbi, but it doesn't seem like a valid connection instance equates to a valid 
> >> and healthy connection. Upon the first attempt to log a message to the db, this 
> >> instance checking works because the connection instance does not yet exist. Once 
> >> the initial connection instance is made, a connection to the database is also 
> >> made (and the first INSERT statement is executed), but the connection instance does 
> >> not guarantee that the connection will remain intact. If the connection is terminated 
> >> outside of syslog-ng, it would seem, the "connection instance" is still valid and 
> >> subsequently a new connection is never attempted and FALSE is never returned. 
> >> 
> >> The table validation check, also in afsql_dd_insert_db(): 
> >> 
> >>    if (!afsql_dd_validate_table(self, table->str))
> >>        {
> >>          /*handle table validation error*/
> >> 
> >> does not catch this either as the table was successfully validated from the first 
> >> successful insert.
> >> 
> >> Keeping inline with the concept that afsql_dd_insert_db() should return FALSE for any
> >> failures, I am proposing the use of the libdbi function dbi_conn_ping() to check if 
> >> the db connection is alive within afsql_dd_insert_db(). Please let me know if I am way
> >> off base here or if I am missing some underlying concepts. I'm sorry that I am not quite 
> >> familiar with git yet, but my proposed patch, with a/modules/afsql/afsql.c being the
> >> current file from syslog-ng 3.2.4, would be as follows:
> >> 
> >> diff -u a/modules/afsql/afsql.c b/modules/afsql/afsql.c 
> >> --- a/modules/afsql/afsql.c	
> >> +++ b/modules/afsql/afsql.c	
> >> @@ -572,7 +572,7 @@
> >>   gint i;
> >>   LogPathOptions path_options = LOG_PATH_OPTIONS_INIT;
> >> 
> >> -  if (!self->dbi_ctx)
> >> +  if (!self->dbi_ctx || dbi_conn_ping(self->dbi_ctx) != 1)
> >>     {
> >>       self->dbi_ctx = dbi_conn_new(self->type);
> >>       if (self->dbi_ctx)
> > 
> > This would probably slow down the insert rate a lot as it'd require a
> > roundtrip to the SQL server before each INSERT.
> > 
> > What I would be interested in why syslog-ng doesn't reestablish the SQL
> > connection at the first error. It should do so.
> > 
> > If an error happens that doesn't mean that the last message is lost. It
> > is retried for 3 times (by default).
> > 
> > -- 
> > Bazsi
> > 
> > 
> 
> In my test environment, syslog-ng does not reestablish the SQL connection and 
> this behavior is very easy to reproduce. I was able to replicate this behavior on 
> a clean install of CentOS 5.7 using syslog-ng OSE 3.2.4 (linux glibc2.3.6 amd64 
> (server) and linux glibc2.3.6 i386 (server)) and MySQL 5.0 (default in CentOS 
> base repo, but shouldn't matter). Once I start logging anything to MySQL, all I 
> have to do is to restart the mysqld process and syslog-ng will start dropping log 
> messages (maxing out attempts). The only explanation I can find as to the cause 
> of this condition is what I outlined in my previous email.
> 
> When this condition is entered, all query attempts will fail despite the  
> MAX_FAILED_ATTEMPTS  value. The function, afsql_dd_insert_db() is not 
> designed to return FALSE when an insert statement fails to execute, but instead, 
> a number of attempts are made using the same connection instance regardless 
> of whether or not syslog-ng is actually connected to the db. This means the 
> connection instance is never released and the thread never suspended for a 
> future attempt at connecting.

Ah, I understand.

> 
> Maybe it makes more sense to implement dbi_conn_ping() after the first 
> unsuccessful query attempt in afsql_dd_insert_db(). That way the extra work 
> needed to check if the connection is alive is only performed if the first attempt of 
> the query fails. The only wasted round trip will occur when at least one failed 
> attempt of the query has been made and the connection is actually alive. 
> dbi_conn_ping() has some interesting functionality (advantages?) where, if 
> certain libdbi driver options are set, a reconnect to the db can occur automatically 
> if the connection is found to be down (obviously this functionality is driver specific 
> as dbi_conn_ping() simply invokes the driver specific ping function). This might 
> save a thread from unnecessary suspension in the future.

I like this one, could you please post a patch to implement it?
Thanks.

> 
> Alternatively, after detecting a failed insert statement, afsql_dd_insert_db() could 
> return FALSE (after incrementing the attempt counter). This will maintain the 
> query attempt count for the duration of subsequent attempts while still allowing 
> for the normal behavior of database connection attempts if, in fact, the db is 
> unreachable for a period of time. This approach would seem to involve the most 
> risk as it changes the way that insert statement errors are handled, while the 
> former only changes behavior if the connection is found to be down after a normal 
> insert error.

-- 
Bazsi