Percona Server crash-safe replication conflicts with WSREP crash recovery

[Raghavendra D Prabhu]

This should only affect a PXC node which is a slave to an async master.

This is because:

#ifdef WITH_WSREP
/* We hijack TRX_SYS_MYSQL_MASTER_LOG_INFO, it seems to be completely unused
   otherwise (see comments for MySQL bug #34058). */
/** */
#define TRX_SYS_WSREP_XID_INFO TRX_SYS_MYSQL_MASTER_LOG_INFO

in trx0sys.h

TRX_SYS_MYSQL_MASTER_LOG_INFO is also used in slave recovery in
PS code.

[Raghavendra D Prabhu (raghavendra-prabhu)]

The quickest solution to this is to add a '} else {' so that they
are mutually exclusive with priority given to wsrep recovery.

[Raghavendra D Prabhu (raghavendra-prabhu)]

@codership, Note that this affects MariaDB Galera cluster too since they use xtradb as well.

#define TRX_SYS_MYSQL_MASTER_LOG_INFO (UNIV_PAGE_SIZE - 2000)
#define TRX_SYS_MYSQL_RELAY_LOG_INFO (UNIV_PAGE_SIZE - 1500)
#define TRX_SYS_COMMIT_MASTER_LOG_INFO (UNIV_PAGE_SIZE - 3000)
#define TRX_SYS_COMMIT_RELAY_LOG_INFO (UNIV_PAGE_SIZE - 2500)
#define TRX_SYS_MYSQL_LOG_INFO (UNIV_PAGE_SIZE - 1000)
#define TRX_SYS_WSREP_XID_INFO TRX_SYS_MYSQL_MASTER_LOG_INFO
#define TRX_SYS_WSREP_XID_LEN (4 + 4 + 4 + XIDDATASIZE)
#define TRX_SYS_DOUBLEWRITE (UNIV_PAGE_SIZE - 200)
#define TRX_SYS_FILE_FORMAT_TAG (UNIV_PAGE_SIZE - 16)

Is it possible to not use any other offset at all? Looks like there is a big gap between UNIV_PAGE_SIZE - 200 and UNIV_PAGE_SIZE 1000 for XID of length 140.

[Raghavendra D Prabhu (raghavendra-prabhu)]

As discussed in the meeting, I will check on how innodb_overwrite_relay_log_info affects this.

[Raghavendra D Prabhu (raghavendra-prabhu)]

OK, I checked but the conflict happens irrespective of the value
of innodb_overwrite_relay_log_info.

For interested, look at trx0trx.c for details.
(innodb_overwrite_relay_log_info works in ha_innodb.cc).

[Ovais Tariq (ovais-tariq)]

So additional work is done in XtraDB irrespective of innodb_overwrite_relay_log_info being enabled or not ?

[Raghavendra D Prabhu (raghavendra-prabhu)]

@Ovais, yes, you can look at trx_write_serialisation_history for details.

[Raghavendra D Prabhu (raghavendra-prabhu)]

From trx_sys.rb:

# The basic structure of a TRX_SYS page is: FIL header, TRX_SYS header,
# empty space, master binary log information, empty space, local binary
# log information, empty space, doublewrite information (repeated twice),
# empty space, and FIL trailer.

So, there is plenty of empty space between TRX_SYS header, as suggested earlier UNIV_PAGE_SIZE - 3500 should be good enough. (for total XID size of about 150 bytes)

[Shahriyar Rzayev (rzayev-sehriyar)]

Percona now uses JIRA for bug reports so this bug report is migrated to: https://jira.percona.com/browse/PXC-965