Stephen O'Donnell created HDDS-6975:
---------------------------------------
Summary: EC: Define the value of Maintenance Redundancy for EC
containers
Key: HDDS-6975
URL: https://issues.apache.org/jira/browse/HDDS-6975
Project: Apache Ozone
Issue Type: Sub-task
Components: SCM
Reporter: Stephen O'Donnell
For Ratis containers, we have a setting
hdds.scm.replication.maintenance.replica.minimum, which defaults to 2. This
indicates how many replicas must still be online when a node is allowed to go
into maintenance.
For EC, we need to decide if we reuse this setting, and what it means.
For example, for Ratis containers, with 3 replicas and the default of 2, you
should be able to take any single node offline without any replication.
With EC, if you had 3-2 containers, and must have a remaining redundancy of 2,
then you must replicate if any node goes offline.
However for the other EC schemes, 6-3, 10-4, they are at least as good as Ratis
with a default of 2.
A better setting for EC might be parityNum / 2, rounded down by integer
division:
3-2 = 1
6-3 = 1
10-4 = 2 remaining redundancy.
Or perhaps parityNum - X, where X is the number of replicas allowed to be
offline in a nEC group.
Its also worth noting that for EC, maintenance mode makes reads much more
expensive. Potentially all reads will turn into online-reconstruction. For
Ratis, it just reduces the available nodes to read from.
With that in mind, another argument for EC, is that all data containers are
kept online for maintenance, with only parity + redundancy allowed to be
offline. I feel that would be a more tricky feature and something we may
consider in the future.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
