This section describes the possible scenarios of a single ESXi host failure and how VMware vSphere's DRS and HA features ensure automatic failover of master hosts and segment hosts to other running ESXi hosts.
How DRS and HA protect a running Greenplum cluster
In a VMware Greenplum on vSphere deployment there is a one-to-one mapping between the Greenplum segments and the segment hosts that are hosting the segments. This means that each segment hosts is only hosting one instance of a Greenplum segment: primary or mirror.
Given the segment hosts host name, you can determine which segment is running in it, as well as what segment hosts are in a primary/mirror pair for a given content id.
For a Greenplum Cluster of 66 hosts deployment, 64 of them are segments named sdw1 to sdw64. There is also the master, mdw, and a standby master smdw.
Use the formula below to compute the content id for each segment. It uses the host name index from the segment hosts: for example, for sdw1 the index is 1.
content_id = floor((hostname_index - 1) / 2)
By applying the formula above, the segment hosts sdw1 and sdw2 map to content ID 0, while sdw3 and sdw4 map to content ID 1, and so on.
The index number lets you determine whether a segment host the primary or the mirror for that content: an odd index is created as a primary, and an even index is created as a mirror. Note that the roles may change when a failover happens (see Best Practices for more details).
During deployment, Terraform created DRS Anti-Affinity rules to ensure that two segment hosts with the same content id never run on the same ESXi host.
When an ESXi host hosting a primary segment host fails, Greenplum always fails over to its corresponding mirror that is safely running on another host.
In the section Setting Up VMware vSphere HA and DRS, you enabled VMware vSphere HA with the following options:
- Host Failure Response/Failure Resource: Restart VMs
- Admission Control/Host failures cluster tolerates: 1
As a result, VMware vSphere HA is configured to tolerate the failure of a single host. This configuration signals the VMware vSphere cluster to reserve the necessary resources, such as CPU and memory, in order to accommodate all the master hosts and segment hosts that were running on the failed host. For example, for a VMware vSphere cluster with 4 hosts, 25% of resources are reserved.
Single Host Failure Scenarios
When an ESXi host goes down, VMware vSphere HA will restart the master hosts and segment hosts that were running on it on other ESXi hosts. There are 4 different scenarios:
Note that since multiple segment hosts run on the same ESXi host, you will encounter more than one of these scenarios during a single host outage. However, you will never experience standby master and master failure for the same host.
Mirror Segment Failure
In this scenario, you will not notice when a mirror segment goes down. It is still possible to perform database operations such as query a table, load data into a table, or create more tables. The detection of a mirror going down can take up to 2 minutes (see Grace Period for details).
However, Greenplum high availability is affected because the mirror cannot replicate the work of the primary.
You can query the gp_segment_configuration table and obtain similar information:
gpadmin=# SELECT * FROM gp_segment_configuration WHERE status = 'd' AND role = 'm';
dbid | content | role | preferred_role | mode | status | port | hostname | address | datadir
------+---------+------+----------------+------+--------+------+----------+---------+--------------------------------
3 | 0 | m | m | n | d | 7000 | sdw2 | sdw2 | /gpdata/primary/gpseg0
The failed mirror segment host will be automatically restarted on another host by VMware vSphere HA. You may run gprecoverseg from the Greenplum master and the mirror segment will be brought back up.
Primary Segment Failure
In this scenario, when a primary segment goes down, all queries running on that segment will be cancelled. You will observe the following error message when trying to run a query:
gpadmin=# SELECT COUNT(DISTINCT c) FROM t t1;
ERROR: FTS detected connection lost during dispatch to seg0 slice1 192.168.11.1:6000 pid=3259:
The Greenplum Fault Tolerance Server (FTS) process will detect the primary segment failing and the failover to its mirror will commence. You will see the role switch on gp_segment_configuration table:
gpadmin=# SELECT * FROM gp_segment_configuration WHERE content = 0;
dbid | content | role | preferred_role | mode | status | port | hostname | address | datadir
------+---------+------+----------------+------+--------+------+----------+---------+--------------------------------------------------
2 | 0 | m | p | n | d | 6000 | sdw1 | sdw1 | /gpdata
3 | 0 | p | m | n | u | 7000 | sdw2 | sdw2 | /gpdata
(2 rows)
The original primary segment will be marked as a mirror and its status will be marked as down.
The original mirror segment will be promoted to be the new primary and, when retrying the failed query above, it will finish successfully:
gpadmin=# SELECT count(DISTINCT c) FROM t t1;
count
-------
101
(1 row)
The failed primary segment host will be automatically restarted by VMware vSphere HA on another host. You may now log in to the Greenplum master host and run gprecoverseg, similar to the step above, to bring back the downed mirror.
Standby Master Failure
When the standby master goes down, the only feature affected by this failure is the reduced availability of the master. Once the standby master has been recovered, it will catch up with the catalog changes on the master.
Before the standby master failure, the gp_segment_configuration table would show the following:
gpadmin=# SELECT * FROM gp_segment_configuration WHERE content = -1;
dbid | content | role | preferred_role | mode | status | port | hostname | address | datadir
------+---------+------+----------------+------+--------+------+----------+---------+--------------------------------------------------
1 | -1 | p | p | n | u | 5432 | mdw | mdw | /gpdata
66 | -1 | m | m | s | u | 5432 | smdw | smdw | /gpdata
(2 rows)
The above table does not reflect state of the standby master. You must query the pg_stat_replication table in order to check the status of the standby master. If the replication process gp_walreceiver is running, it means that the standby master is up.
gpadmin=# SELECT usename, application_name, client_addr FROM pg_stat_replication;
usename | application_name | client_addr
---------+------------------+----------------
gpadmin | gp_walreceiver | 192.168.11.251
(1 row)
When the standby master is down, the output of gpstate will return output similar to the following:
[gpadmin@mdw ~]$ gpstate
<Date-Time> gpstate:mdw:gpadmin-[INFO]:-Greenplum instance status summary
<Date-Time> gpstate:mdw:gpadmin-[INFO]:-----------------------------------------------------
<Date-Time> gpstate:mdw:gpadmin-[INFO]:- Master instance = Active
<Date-Time> gpstate:mdw:gpadmin-[INFO]:- Master standby = smdw
<Date-Time> gpstate:mdw:gpadmin-[WARNING]:-Standby master state = Standby host DOWN <<<<<<<<
<Date-Time> gpstate:mdw:gpadmin-[INFO]:- Total segment instance count from metadata = 64
<Date-Time> gpstate:mdw:gpadmin-[INFO]:-----------------------------------------------------
If the standby master host goes down, VMware vSphere HA will automatically restart it on another host.
You must run gpinitstandby -n from the Greenplum master to bring the standby master back up.
[gpadmin@mdw ~]$ gpinitstandby -n
<Date-Time> gpinitstandby:mdw:gpadmin-[INFO]:-Starting standby master
<Date-Time> gpinitstandby:mdw:gpadmin-[INFO]:-Checking if standby master is running on host: smdw in directory: /gpdata
<Date-Time> gpinitstandby:mdw:gpadmin-[INFO]:-Successfully started standby master
You may verify that the standby master is up and running by running gpstate:
[gpadmin@mdw ~]$ gpstate
<Date-Time> gpstate:mdw:gpadmin-[INFO]:-Greenplum instance status summary
<Date-Time> gpstate:mdw:gpadmin-[INFO]:-----------------------------------------------------
<Date-Time> gpstate:mdw:gpadmin-[INFO]:- Master instance = Active
<Date-Time> gpstate:mdw:gpadmin-[INFO]:- Master standby = smdw
<Date-Time> gpstate:mdw:gpadmin-[INFO]:- Standby master state = Standby host passive
<Date-Time> gpstate:mdw:gpadmin-[INFO]:- Total segment instance count from metadata = 64
<Date-Time> gpstate:mdw:gpadmin-[INFO]:-----------------------------------------------------
Master Failure
If the Greenplum master goes down, you will lose connection to the entire Greenplum cluster. This is a highly disruptive experience because all client connections will be dropped.
VMware vSphere HA will automatically restart the failed master host on another host. Once the master host has been restarted, you must run gpstart to restart the cluster.
Note that this scenario assumes that Greenplum master host comes back up without issues when starting it on a different host. In the event of Greenplum master corruption, see Recovering a Failed Master in order to activate the standby master.
After Host Recovery
When a failed ESXi host comes back up, because the failed master hosts and segment hosts have already been restarted on other running hosts by VMware vSphere HA, VMware vSphere DRS will rearrange the master hosts and segment hosts, not necessarily the ones that went down, to balance the load of all the ESXi hosts. This operation is done through VMware vSphere vMotion, which is transparent to the running master hosts and segment hosts.
There is no need for an explicit rebalancing operation because the segment hosts are all identical by design, primaries and mirrors. Service disruption occurs only once during host failure, and only in the case of a primary or master failure.
Best Practices for Greenplum HA
In most cases, running gprecoverseg will recover the failed segments with no issues when the affected segment host is restarted by VMware vSphere HA on a different ESXi host. VMware recommends you monitor the status of the cluster in order to run gprecoverseg when required. Refer to Monitoring tasks from the Greenplum Administrator Guide. Alternatively, you may use Greenplum Command Center to set up alerts.
After a segment recovery, segment instances may not be returned to the preferred role that they were given at system initialization time. This will have no effect on the performance for a Greenplum on VMware vSphere configuration.
However, you may want to maintain consistency between the segment host host names and their preferred role (odd numbers for primaries, even numbers for mirrors) for monitoring purposes. Consider running gprecoverseg -r periodically to rebalance segment roles after segment failures. Note that queries in progress will be canceled and rolled back, so you may want to schedule this operation during a maintenance window.
Appendix: Grace Period for a Downed Segment
There are two Greenplum configuration parameters that control the detection of a downed mirror:
gp_fts_mark_mirror_down_grace_periodcontrols the waiting time between primary and its mirror. The default value is 30 seconds.gp_fts_probe_intervalcontrols the FTS probe interval from the master to the primaries. The default value is 60 seconds.
Combining the values of both parameters, it will take up to two minutes to detect a mirror segment going down.
Note: If both the primary and mirror segment hosts with the same content id are restarted during that period, Greenplum Database cluster will hit a double fault, during which it is no longer able to process queries. In this case, the only way to bring the cluster back online is to restart the entire Greenplum cluster using gpstop -r.
