m4_comment([$Id: rep_elect.so,v 1.26 2007/03/20 20:20:26 alanb Exp $])

define(M4PAGELOCAL, [rep_elect, DB_REP_UNAVAIL])
include(m4/m4.seealso)

m4_pf_header(m4_ref(rep_elect),
ifelse(M4API, C_API, [dnl
int
DB_ENV-__GT__rep_elect(DB_ENV *env,
    int nsites, int nvotes, u_int32_t flags);
])
ifelse(M4API, CXX_API, [dnl
int
DbEnv::rep_elect(int nsites, int nvotes, u_int32_t flags);
]))

m4_p([dnl
The m4_refT(rep_elect) holds an election for the master of a replication
group.])

m4_repl_lower(rep_elect)

m4_p([dnl
If the election is successful, m4_db will notify the application of
the results of the election by means of either the
m4_ref(DB_EVENT_REP_ELECTED) or m4_ref(DB_EVENT_REP_NEWMASTER) events
(see m4_refT(dbenv_set_event_notify) for more information).  The
application is responsible for adjusting its relationship to the other
database environments in the replication group, including directing
all database updates to the newly selected master, in accordance with
the results of the election.])

m4_p([dnl
The thread of control that calls the m4_refT(rep_elect) must not be the
thread of control that processes incoming messages; processing the
incoming messages is necessary to successfully complete an election.])

m4_parambegin
m4_param(nsites, [dnl
The m4_arg(nsites) parameter specifies the number of replication sites
expected to participate in the election.  Once the current site has
election information from that many sites, it will short-circuit the
election and immediately cast its vote for a new master. The
m4_arg(nsites) parameter must be a positive integer, no less than
m4_arg(nvotes), or 0 if the election should use the value previously set
using the m4_refT(rep_set_nsites).])

m4_param(nvotes, [dnl
The m4_arg(nvotes) parameter specifies the minimum number of replication
sites from which the current site must have election information, before
the current site will cast a vote for a new master.  The m4_arg(nvotes)
parameter must be a positive integer and no greater than m4_arg(nsites),
or 0 if the election should use the value ((m4_arg(nsites) / 2) + 1) as
the m4_arg(nvotes) argument.])

m4_unusedflags
m4_paramend

m4_p([dnl
Elections are done in two parts: first, replication sites collect
information from the other replication sites they know about, and
second, replication sites cast their votes for a new master.  The second
phase is triggered by one of two things: either the replication site
gets election information from m4_arg(nsites) sites, or the election
m4_arg(timeout) expires.  Once the second phase is triggered, the
replication site will cast a vote for the new master of its choice if,
and only if, the site has election information from at least
m4_arg(nvotes) sites.  If a site receives m4_arg(nvotes) votes for it
to become the new master, then it will become the new master.])

m4_p([dnl
We recommend m4_arg(nvotes) be set to at least:])
m4_indent([(sites participating in the election / 2) + 1])
m4_p([dnl
to ensure there are never more than two masters active at the same time
even in the case of a network partition.  When a network partitions, the
side of the partition with more than half the environments will elect a
new master and continue, while the environments communicating with fewer
than half of the environments will fail to find a new master, as no site
can get m4_arg(nvotes) votes.])

m4_p([dnl
We recommend m4_arg(nsites) be set to:])
m4_indent([number of sites in the replication group - 1])
m4_p([dnl
when choosing a new master after a current master fails.  This allows
the group to reach a consensus without having to wait for the timeout
to expire.])

m4_p([dnl
When choosing a master from among a group of client
sites all restarting at the same time, it makes more sense to set
m4_arg(nsites) to the total number of sites in the group, since there is no
known missing site.  Furthermore, in order to ensure the best choice
from among sites that may take longer to boot than the local site,
setting m4_arg(nvotes) also to this same total number of sites will
guarantee that every site in the group is considered.  Alternatively,
using the special timeout for full elections allows full participation
on restart but allows election of a master if one site does not
reboot and rejoin the group in a reasonable amount of time. (See the
m4_link(M4RELDIR/ref/rep/elect, [Elections])
section in the m4_db Reference Guide for more information.)])

m4_p([dnl
Setting m4_arg(nsites) to lower values can increase the speed of an
election, but can also result in election failure, and is usually not
recommended.])

m4_idefz(DB_REP_UNAVAIL)
m4_err(rep_elect,
DB_REP_UNAVAIL, [dnl
The replication group was unable to elect a master, or was unable to
complete the election in the election timeout period (see
m4_refT(rep_set_timeout) for more information).])

m4_seealso(Replication)
m4_page_footer