Files
repmgr/docs/repmgrd-failover-mechanism.md
2016-10-05 11:25:36 +09:00

76 lines
3.7 KiB
Markdown

repmgrd's failover algorithm
============================
When implementing automatic failover, there are two factors which are critical in
ensuring the desired result is achieved:
- has the master node genuinely failed?
- which is the best node to promote to the new master?
This document outlines repmgrd's decision-making process during automatic failover
for standbys directly connected to the master node.
Master node failure detection
-----------------------------
If a `repmgrd` instance running on a PostgreSQL standby node is unable to connect to
the master node, this doesn't neccesarily mean that the master is down and a
failover is required. Factors such as network connectivity issues could mean that
even though the standby node is isolated, the replication cluster as a whole
is functioning correctly, and promoting the standby without further verification
could result in a "split-brain" situation.
In the event that `repmgrd` is unable to connect to the master node, it will attempt
to reconnect to the master server several times (as defined by the `reconnect_attempts`
parameter in `repmgr.conf`), with reconnection attempts occuring at the interval
specified by `reconnect_interval`. This happens to verify that the master is definitively
not accessible (e.g. that connection was not lost due to a brief network glitch).
Appropriate values for these settings will depend very much on the replication
cluster environment. There will necessarily be a trade-off between the time it
takes to assume the master is not reachable, and the reliability of that conclusion.
A standby in a different physical location to the master will probably need a longer
check interval to rule out possible network issues, whereas one located in the same
rack with a direct connection between servers could perform the check very quickly.
Note that it's possible the master comes back online after this point is reached,
but before a new master has been selected; in this case it will be noticed
during the selection of a new master and no actual failover will take place.
Promotion candidate selection
-----------------------------
Once `repmgrd` has decided the master is definitively unreachable, following checks
will be carried out:
* attempts to connect to all other nodes in the cluster (including the witness
node, if defined) to establish the state of the cluster, including their
current LSN
* If less than half of the nodes are visible (from the viewpoint
of this node), `repmgrd` will not take any further action. This is to ensure that
e.g. if a replication cluster is spread over multiple data centres, a split-brain
situation does not occur if there is a network failure between datacentres. Note
that if nodes are split evenly between data centres, a witness server can be
used to establish the "majority" daat centre.
* `repmgrd` polls all visible servers and waits for each node to return a valid LSN;
it updates the LSN previously stored for this node if it has increased since
the initial check
* once all LSNs have been retrieved, `repmgrd` will check for the highest LSN; if
its own node has the highest LSN, it will attempt to promote itself (using the
command defined in `promote_command` in `repmgr.conf`. Note that if using
`repmgr standby promote` as the promotion command, and the original master becomes available
before the promotion takes effect, `repmgr` will return an error and no promotion
will take place, and `repmgrd` will resume monitoring as usual.
* if the node is not the promotion candidate, `repmgrd` will execute the
`follow_command` defined in `repmgr.conf`. If using `repmgr standby follow` here,
`repmgr` will attempt to detect the new master node and attach to that.