[
https://issues.apache.org/jira/browse/CASSANDRA-4583?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Edward Sargisson updated CASSANDRA-4583:
Description:
At present we do not have a complete reproduction for this defect but am
raising this defect as request by Aaron Morton. We will update as we find out
more. If any additional logging or tests are requested we will do them if we
can.
We have experienced 2 failures ascribed to this defect. On the cassandra user
mailing list Peter Schuller (2012-08-28) describes an additional failure.
Reproduction steps as currently known:
1. Setup a cluster with 6 nodes (call them #1 through #6).
2. Have #5 fail completely. One failure was when the node was stopped to
replace the battery in the hard disk cache. The second failure was when the
hardware monitoring recorded a problem, CPU usage was increasing without
explanation and the server console was frozen so the machine was restarted.
3. Bring #5 back
Expected behaviour:
* #5 should rejoin the ring.
Actual behaviour (based on the incident we saw yesterday):
* #5 didn't rejoin the ring.
* We stopped all nodes and started them one by one.
* Nodes #2, #4, #6 had forgotten most of their column families. They had the
keys space but with only one column family instead of the usual 9 or so.
* We ran nodetool resetlocalschema on #2, #4 and #6.
* We ran nodetool repair -pr on #2, #4, #5 and #6
* On #2 nodetool repair appeared to crash in that there were no messages in the
logs from it for 10min+. Nodetool compactionstats and nodetool netstats showed
no activity.
* Restarting nodetool repair -pr fixed the problem and ran to completion.
was:
At present we do not have a complete reproduction for this defect but am
raising this defect as request by Aaron Morton. We will update as we find out
more. If any additional logging or tests are requested we will do them if we
can.
We have experienced 2 failures ascribed to this defect. On the cassandra user
mailing list Peter Schuller (2012-08-28) describes an additional failure.
Reproduction steps as currently known:
1. Setup a cluster with 6 nodes (call them #1 through #6).
2. Have #5 fail completely. One failure was when the node was stopped to
replace the battery in the hard disk cache. The second failure was when the
hardware monitoring recorded a problem, CPU usage was increasing without
explanation and the server console was frozen so the machine was restarted.
3. Bring #5 back
Expected behaviour:
* #5 should rejoin the ring.
Actual behaviour (based on the incident we saw yesterday):
* #5 didn't rejoin the ring.
* We stopped all nodes and started them one by one.
* Nodes #2, #4, #6 had forgotten most of their column families. They had the
keys space but with only one column family instead of the usual 9 or so.
* We ran nodetool resetlocalschema on #2, #4 and #6.
* We ran nodetool repair -pr on #2, #4, #5 and #6
* On one of these nodes nodetool repair appeared to crash in that there were no
messages in the logs from it for 10min+. Nodetool compactionstats and nodetool
netstats showed no activity.
* Restarting nodetool repair -pr fixed the problem and ran to completion.
Some nodes forget schema when 1 node fails
--
Key: CASSANDRA-4583
URL: https://issues.apache.org/jira/browse/CASSANDRA-4583
Project: Cassandra
Issue Type: Bug
Components: Core
Affects Versions: 1.1.2
Environment: CentOS release 6.3 (Final)
Reporter: Edward Sargisson
At present we do not have a complete reproduction for this defect but am
raising this defect as request by Aaron Morton. We will update as we find out
more. If any additional logging or tests are requested we will do them if we
can.
We have experienced 2 failures ascribed to this defect. On the cassandra user
mailing list Peter Schuller (2012-08-28) describes an additional failure.
Reproduction steps as currently known:
1. Setup a cluster with 6 nodes (call them #1 through #6).
2. Have #5 fail completely. One failure was when the node was stopped to
replace the battery in the hard disk cache. The second failure was when the
hardware monitoring recorded a problem, CPU usage was increasing without
explanation and the server console was frozen so the machine was restarted.
3. Bring #5 back
Expected behaviour:
* #5 should rejoin the ring.
Actual behaviour (based on the incident we saw yesterday):
* #5 didn't rejoin the ring.
* We stopped all nodes and started them one by one.
* Nodes #2, #4, #6 had forgotten most of their column families. They had the
keys space but with only one column family instead of the usual 9 or so.
* We ran nodetool resetlocalschema on #2, #4 and #6.
* We ran nodetool repair -pr on #2, #4, #5 and #6
* On #2 nodetool repair appeared to crash in that