Mathieu Roy created CASSANDRA-10501:
---------------------------------------
Summary: Failure to start up Cassandra when temporary compaction
files are not all renamed after kill/crash (FSReadError)
Key: CASSANDRA-10501
URL: https://issues.apache.org/jira/browse/CASSANDRA-10501
Project: Cassandra
Issue Type: Bug
Components: Core
Environment: Cassandra 2.1.6
Redhat Linux
Reporter: Mathieu Roy
We have seen an issue intermittently but repeatedly over the last few months
where, after exiting the Cassandra process, it fails to start with an
FSReadError (stack trace below). The FSReadError refers to a 'statistics' file
for a that doesn't exist, though a corresponding temporary file does exist
(eg. there is no
/media/data/cassandraDB/data/clusteradmin/singleton_token-01a92ed069b511e59b2c53679a538c14/clusteradmin-singleton_token-ka-9-Statistics.db
file, but there is a
/media/data/cassandraDB/data/clusteradmin/singleton_token-01a92ed069b511e59b2c53679a538c14/clusteradmin-singleton_token-tmp-ka-9-Statistics.db
file.)
We tracked down the issue to the fact that the process exited with leftover
compactions and some of the 'tmp' files for the SSTable had been renamed to
final files, but not all of them - the issue happens if the 'Statistics' file
is not renamed but others are. The scenario we've seen on the last two
occurrences involves the 'CompressionInfo' file being a final file while all
other files for the SSTable generation were left with 'tmp' names.
When this occurs, Cassandra cannot start until the file issue is resolved;
we've worked around it by deleting the SSTable files from the same generation,
both final and tmp, which at least allows Cassandra to start. Renaming all
files to either tmp or final names would also work.
We've done some debugging in Cassandra and have been unable to cause the issue
without renaming the files manually. The rename code at SSTableWriter.rename()
looks like it could result in this if the process exits in the middle of the
rename, but in every occurrence we've debugged through, the Set of components
is ordered and Statistics is the first file renamed.
However the comments in SSTableWriter.rename() suggest that the 'Data' file is
meant to be used as meaning the files were completely renamed. The method
ColumnFamilyStore. removeUnfinishedCompactionLeftovers(), however, will proceed
assuming the compaction is complete if any of the component files has a final
name, and will skip temporary files when reading the list. If the 'Statistics'
file is temporary then it won't be read, and the defaults does not include a
list of ancestors, leading to the NullPointerException.
It appears that ColumnFamilyStore. removeUnfinishedCompactionLeftovers() should
perhaps either ensure that all 'tmp' files are properly renamed before it uses
them, or skip SSTable files that don't have either the 'Data' or 'Statistics'
file in final form.
Stack trace:
FSReadError in Failed to remove unfinished compaction leftovers (file:
/media/data/cassandraDB/data/clusteradmin/singleton_token-01a92ed069b511e59b2c53679a538c14/clusteradmin-singleton_token-ka-9-Statistics.db).
See log for details.
at
org.apache.cassandra.db.ColumnFamilyStore.removeUnfinishedCompactionLeftovers(ColumnFamilyStore.java:617)
at
org.apache.cassandra.service.CassandraDaemon.setup(CassandraDaemon.java:302)
at
org.apache.cassandra.service.CassandraDaemon.activate(CassandraDaemon.java:536)
at
org.apache.cassandra.service.CassandraDaemon.main(CassandraDaemon.java:625)
Caused by: java.lang.NullPointerException
at
org.apache.cassandra.db.ColumnFamilyStore.removeUnfinishedCompactionLeftovers(ColumnFamilyStore.java:609)
... 3 more
Exception encountered during startup: java.lang.NullPointerException
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
