[
https://issues.apache.org/jira/browse/KUDU-888?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Mike Percy updated KUDU-888:
----------------------------
Description:
Specifying and validating a remote peer's UUID during communication is required
for correctness, because we assume that if a peer has forgotten or lost its
durable consensus state then it has also forgotten or lost its UUID, so it will
get a new one assigned. Violations of the consensus protocol can occur if a
peer has amnesia, but if it refuses requests intended for a prior incarnation
of itself then it should be possible to maintain safety.
See more discussion of this amnesia scenario in the remote bootstrap design
doc, under "Failure of a disk with consensus metadata is a catastrophic
failure":
https://github.com/apache/kudu/blob/master/docs/design-docs/raft-tablet-copy.md
This also helps guard against bad misconfigurations, such as a tablet sending a
DeleteTablet RPC message to itself when it thinks it is deleting a remote
replica.
was:
Specifying and validating a remote peer's UUID during communication is required
for correctness, because we assume that if a peer has forgotten or lost its
durable consensus state then it has also forgotten or lost its UUID, so it will
get a new one assigned. Violations of the consensus protocol can occur if a
peer has amnesia, but if it refuses requests intended for a prior incarnation
of itself then it should be possible to maintain safety.
See more discussion of this amnesia scenario in the remote bootstrap design
doc, under "Failure of a disk with consensus metadata is a catastrophic
failure":
https://docs.google.com/document/d/1zSibYnwPv9cFRnWn0ORyu2uCGB9Neb-EsF0M6AiMSEE/edit#heading=h.s6ennwonsuz8
This also helps guard against bad misconfigurations, such as a tablet sending a
DeleteTablet RPC message to itself when it thinks it is deleting a remote
replica.
> Consensus peers must specify and validate each others' UUIDs when
> communicating
> -------------------------------------------------------------------------------
>
> Key: KUDU-888
> URL: https://issues.apache.org/jira/browse/KUDU-888
> Project: Kudu
> Issue Type: Bug
> Components: consensus
> Affects Versions: Public beta
> Reporter: Mike Percy
> Assignee: Mike Percy
> Priority: Critical
>
> Specifying and validating a remote peer's UUID during communication is
> required for correctness, because we assume that if a peer has forgotten or
> lost its durable consensus state then it has also forgotten or lost its UUID,
> so it will get a new one assigned. Violations of the consensus protocol can
> occur if a peer has amnesia, but if it refuses requests intended for a prior
> incarnation of itself then it should be possible to maintain safety.
> See more discussion of this amnesia scenario in the remote bootstrap design
> doc, under "Failure of a disk with consensus metadata is a catastrophic
> failure":
> https://github.com/apache/kudu/blob/master/docs/design-docs/raft-tablet-copy.md
> This also helps guard against bad misconfigurations, such as a tablet sending
> a DeleteTablet RPC message to itself when it thinks it is deleting a remote
> replica.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
