Page MenuHomeClusterLabs Projects
Files F2823058

No OneTemporary
Actions

View Options

diff --git a/README b/README
index 74ac934..535df77 100644
--- a/README
+++ b/README
@@ -1,98 +1,159 @@
The Booth Cluster Ticket Manager
=============
Booth manages the ticket which authorizes one of the cluster
sites located in geographically dispersed distances to run
certain resources. It is designed to be an add-on of Pacemaker,
which extends Pacemaker to support geographically distributed
clustering.
Booth is based on the Raft consensus algorithm. Though the
implementation is not complete (there is no log), booth
guarantees that the ticket is always available at just one site
as long as it has exclusive control of the tickets.
The git repository is available at github:
<https://github.com/ClusterLabs/booth>
github can also track issues or bug reports.
Description of a booth cluster
==============================
Booth cluster is a collection of cooperative servers
communicating using the booth protocol. The purpose of the booth
cluster is to manage cluster tickets. The booth cluster consists
of at least three servers.
A booth server can be either a site or an arbitrator. Arbitrators
-help resolve ties in elections.
+help resolve ties in elections, but cannot hold tickets.
The basic unit of operation in the booth cluster is a ticket.
Every non-granted ticket is in the initial state on all servers.
For granted tickets, the server holding the ticket is the leader
and other servers are followers. The leader issues heartbeats and
ticket updates to the followers. The followers are required to
obey the leader.
Booth startup
------------
On startup, the booth process first loads tickets, if available,
from the CIB. Afterwards, it broadcasts a query to get tickets'
status from other servers. Local copies are updated from ticket
replies.
If the server discovers that itself is the ticket leader, it
tries to establish its authority again by broadcasting heartbeat.
If it succeeds, it continues as the leader for this ticket. The
other booth servers become followers. This procedure is possible
only immediately after the booth startup.
Grant and revoke operations
------------
A ticket first has to be granted using the 'booth client grant'
command.
Obviously, it is not possible to grant a ticket which is
currently granted.
Ticket revoke is the operation which is the opposite of grant.
An administrative revoke may be started at any server, but the
operation itself happens only at the leader. If the leader is
unreachable, the ticket cannot be revoked.
A ticket grant may be delayed if not all sites are reachable.
The delay is the ticket expiry time extended by acquire-after, if
set. This is to ensure that the unreachable site relinquished the
ticket it may have been holding and stopped the corresponding
cluster resources.
If the user is absolutely sure that the unreachable site does not
-hold the ticket, the delay may be skipped by using the '--force'
+hold the ticket, the delay may be skipped by using the '-F'
option of the 'booth grant' command.
If in effect, the grant delay time is shown in the 'booth list'
command output.
-Ticket management
+Ticket management and server operation
------------
A granted ticket is managed by the booth servers so that its
availability is maximized without breaking the basic guarantee
that the ticket is granted to one site only.
The server where the ticket is granted is the leader, the other
servers are followers. The leader occasionally sends heartbeats,
once every half ticket expiry under normal circumstances.
If a follower doesn't hear from the leader longer than the ticket
expiry time, it will consider the ticket lost, and try to acquire
it by starting new elections.
-Elections are described in the raft paper at
+A server starts elections by broadcasting the REQ_VOTE RPC.
+Other servers reply with the VOTE_FOR RPC, in which they record
+its vote. Normally, the sender of the first REQ_VOTE gets the
+vote of the receiver. Whichever server gets a majority of votes
+wins the elections. On ties, elections are restarted. To
+decrease chance of elections ending in a tie, a server waits for a
+short random period before sending out the REQ_VOTE packets.
+Everything else being equal, the server which sends REQ_VOTE
+first gets elected.
+
+Elections are described in more detail in the raft paper at
<https://ramcloud.stanford.edu/wiki/download/attachments/11370504/raft.pdf>.
+Ticket renewal (or update) is a two-step process. Before actually
+writing the ticket to the CIB, the server holding the ticket
+first tries to establish that it still has the majority for that
+ticket. That is done by broadcasting a heartbeat. If the server
+receives enough acknowledgements, it then stores the ticket to
+the CIB and broadcasts the UPDATE RPC with updated ticket expiry
+time so that the followers can update local ticket copies. Ticket
+renewals run every half ticket expire time.
+
+Before ticket renewal, the leader runs an external program if
+such program is set in 'before-acquire-handler'. The external
+program should ensure that the cluster managed service which is
+protected by this ticket can run at this site. If that program
+fails, the leader relinquishes the ticket. It announces its
+intention to step down by broadcasting an unsolicited VOTE_FOR
+with an empty vote. On receiving such RPC other servers start new
+elections to elect a new leader.
+
+Split brain
+------------
+
+On split brains two possible issues arise: leader in minority and
+follower disconnected from the leader.
+
+Let's take a look at the first one. The leader in minority
+eventually expires the ticket because it cannot receieve majority
+of acknowledgements in reply to its heartbeats. The other
+partition runs elections (at about the same time, as they find
+the ticket lost after its expiry) and, if it can get the
+majority, the elections winner becomes a new leader for the
+ticket. After split brain gets resolved, the old leader will
+become follower as soon as it receives heartbeat from the new
+leader. Note the timing: the old leader releases the ticket at
+around the same time as when new elections in the other partition
+are held. This is because the leader ensures that the ticket
+expire time is always the same on all servers in the booth
+cluster.
+
+The second situation, where a follower is disconnected from the
+leader, is a bit more difficult to handle. After the ticket
+expiry time, the follower will consider the ticket lost and start
+new elections. The elections repeatedly get restarted until the
+split brain is resolved. Then, the rest of the cluster send
+rejects in reply to REQ_VOTE RPC because the ticket is still
+valid and therefore couldn't have been lost. They know that
+because the reason for elections is included with every REQ_VOTE.
+
+Short intermittent split brains are handled well because the
+leader keeps resending heartbeats until it gets replies from all
+servers serving sites.
+
# vim: set ft=asciidoc :

File Metadata

Mime Type
text/x-diff
Expires
Sat, Jan 25, 7:08 AM (1 d, 15 h)
Storage Engine
blob
Storage Format
Raw Data
Storage Handle
1321100
Default Alt Text
(6 KB)

Event Timeline