[jira] [Commented] (CASSANDRA-18980) Bring PropertyFileSnitch & RackInferringSnitch into line with TCM
[ https://issues.apache.org/jira/browse/CASSANDRA-18980?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17786853#comment-17786853 ] Marcus Eriksson commented on CASSANDRA-18980: - +1 > Bring PropertyFileSnitch & RackInferringSnitch into line with TCM > - > > Key: CASSANDRA-18980 > URL: https://issues.apache.org/jira/browse/CASSANDRA-18980 > Project: Cassandra > Issue Type: Task > Components: Transactional Cluster Metadata >Reporter: Sam Tunnicliffe >Assignee: Sam Tunnicliffe >Priority: Normal > Attachments: ci_summary.html > > > With TCM, all snitch implementations should only be responsible for the > initial the topology configuration of the local node. For example, PFS should > behave more like {{GossipingPropertyFileSnitch}} in that its configuration > file specifies the DC & rack for the local node only. Each node in the > cluster is responsible for updating its own topology info in > {{ClusterMetadata}} when it is registered and the location of peers can be > easily looked up there. > Modifying the DC & rack of a live node via GPFS/PFS config is desabled by > default and strongly discouraged > (CASSANDRA-10242/CASSANDRA-10243/CASSANDRA-9474) and there are > {{StartupChecks}} in place to prevent changing location at startup. > Given that modifying location of a joined node is a fundamentally unsafe > operation, we should remove the ability to do so. The argument for allowing > it has previously been that if a DC or cluster has no data, then the overhead > of decommissioning and re-joining nodes to fix location is prohibitive. > CEP-21 fixes that by making those operations more reliable, so the override > becomes less valuable. > Implementation can be found in: > https://github.com/beobal/cassandra/commit/8ec1ea705210b7e2e38d888768ba83faa53754d8 > https://github.com/beobal/cassandra/commit/7b41f063e04f25b8215318f51d908732b1f21f5f > https://github.com/beobal/cassandra/commit/09cdc5351ad049d97c7eeb53c4e68701c088a711 > CI summary results attached. For right now, we've only been foccussing on > unit tests, in-jvm & python dtests, so the failures in other suites are to be > expected. Unfortunately, the archive containing the detailed results is too > large to attach as a single file, so will look into the best way to work > around that. > The intention is to fix all suites before merging the feature branch. > ||Suite||Total||Passed||Skipped||Failed||Errors|| > |jvm dtest|1508|1465|24|8|11| > |python dtest|1083|790|262|28|3| > |unit|12082|12006|61|6|9| -- This message was sent by Atlassian Jira (v8.20.10#820010) - To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org For additional commands, e-mail: commits-h...@cassandra.apache.org
[jira] [Commented] (CASSANDRA-18980) Bring PropertyFileSnitch & RackInferringSnitch into line with TCM
[ https://issues.apache.org/jira/browse/CASSANDRA-18980?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17786855#comment-17786855 ] Alex Petrov commented on CASSANDRA-18980: - +1 > Bring PropertyFileSnitch & RackInferringSnitch into line with TCM > - > > Key: CASSANDRA-18980 > URL: https://issues.apache.org/jira/browse/CASSANDRA-18980 > Project: Cassandra > Issue Type: Task > Components: Transactional Cluster Metadata >Reporter: Sam Tunnicliffe >Assignee: Sam Tunnicliffe >Priority: Normal > Attachments: ci_summary.html > > > With TCM, all snitch implementations should only be responsible for the > initial the topology configuration of the local node. For example, PFS should > behave more like {{GossipingPropertyFileSnitch}} in that its configuration > file specifies the DC & rack for the local node only. Each node in the > cluster is responsible for updating its own topology info in > {{ClusterMetadata}} when it is registered and the location of peers can be > easily looked up there. > Modifying the DC & rack of a live node via GPFS/PFS config is desabled by > default and strongly discouraged > (CASSANDRA-10242/CASSANDRA-10243/CASSANDRA-9474) and there are > {{StartupChecks}} in place to prevent changing location at startup. > Given that modifying location of a joined node is a fundamentally unsafe > operation, we should remove the ability to do so. The argument for allowing > it has previously been that if a DC or cluster has no data, then the overhead > of decommissioning and re-joining nodes to fix location is prohibitive. > CEP-21 fixes that by making those operations more reliable, so the override > becomes less valuable. > Implementation can be found in: > https://github.com/beobal/cassandra/commit/8ec1ea705210b7e2e38d888768ba83faa53754d8 > https://github.com/beobal/cassandra/commit/7b41f063e04f25b8215318f51d908732b1f21f5f > https://github.com/beobal/cassandra/commit/09cdc5351ad049d97c7eeb53c4e68701c088a711 > CI summary results attached. For right now, we've only been foccussing on > unit tests, in-jvm & python dtests, so the failures in other suites are to be > expected. Unfortunately, the archive containing the detailed results is too > large to attach as a single file, so will look into the best way to work > around that. > The intention is to fix all suites before merging the feature branch. > ||Suite||Total||Passed||Skipped||Failed||Errors|| > |jvm dtest|1508|1465|24|8|11| > |python dtest|1083|790|262|28|3| > |unit|12082|12006|61|6|9| -- This message was sent by Atlassian Jira (v8.20.10#820010) - To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org For additional commands, e-mail: commits-h...@cassandra.apache.org