[
https://issues.apache.org/jira/browse/CASSANDRA-9120?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14483344#comment-14483344
]
Ariel Weisberg commented on CASSANDRA-9120:
-------------------------------------------
If anyone who has some of these problem files could make available a copy that
would be helpful. Or just the first few kilobytes. I would like to take a look
at just what kind of problem value we are seeing. The odds of multiple people
seeing corruption of the same length prefix bytes seems a little low.
I feel like trying to figure out memory available and then decide how many keys
to load is not a good heuristic to attempt. For older versions where we don't
have a checksum I think the proposed patch makes sense. That is due to not
having any better options when there is no checksum.
Operators always have the option of removing the problem key caches so
incentive to put together anything that might be fragile (like partially
loading caches) isn't strong IMO.
I do see an issue where an operator has changing memory layout (node size, heap
size) or utilization and would want to be able to reduce the % of keys loaded.
I am guessing the story for that now is that they delete the key cache?
I don't see how # of keys saved matters. The # of keys saved is bounded by the
# of keys loaded (or is it?) which means it fit into memory in the first place
and should do it again on restart.
Before we go ahead an implement something to work around this I would like to
get a better sense of how we are arriving at this failure in the first place.
> OutOfMemoryError when read auto-saved cache (probably broken)
> -------------------------------------------------------------
>
> Key: CASSANDRA-9120
> URL: https://issues.apache.org/jira/browse/CASSANDRA-9120
> Project: Cassandra
> Issue Type: Bug
> Environment: Linux
> Reporter: Vladimir
> Assignee: Jeff Jirsa
> Fix For: 3.0, 2.0.15, 2.1.5
>
>
> Found during tests on a 100 nodes cluster. After restart I found that one
> node constantly crashes with OutOfMemory Exception. I guess that auto-saved
> cache was corrupted and Cassandra can't recognize it. I see that similar
> issues was already fixed (when negative size of some structure was read).
> Does auto-saved cache have checksum? it'd help to reject corrupted cache at
> the very beginning.
> As far as I can see current code still have that problem. Stack trace is:
> {code}
> INFO [main] 2015-03-28 01:04:13,503 AutoSavingCache.java (line 114) reading
> saved cache
> /storage/core/loginsight/cidata/cassandra/saved_caches/system-sstable_activity-KeyCache-b.db
> ERROR [main] 2015-03-28 01:04:14,718 CassandraDaemon.java (line 513)
> Exception encountered during startup
> java.lang.OutOfMemoryError: Java heap space
> at java.util.ArrayList.<init>(Unknown Source)
> at
> org.apache.cassandra.db.RowIndexEntry$Serializer.deserialize(RowIndexEntry.java:120)
> at
> org.apache.cassandra.service.CacheService$KeyCacheSerializer.deserialize(CacheService.java:365)
> at
> org.apache.cassandra.cache.AutoSavingCache.loadSaved(AutoSavingCache.java:119)
> at
> org.apache.cassandra.db.ColumnFamilyStore.<init>(ColumnFamilyStore.java:262)
> at
> org.apache.cassandra.db.ColumnFamilyStore.createColumnFamilyStore(ColumnFamilyStore.java:421)
> at
> org.apache.cassandra.db.ColumnFamilyStore.createColumnFamilyStore(ColumnFamilyStore.java:392)
> at org.apache.cassandra.db.Keyspace.initCf(Keyspace.java:315)
> at org.apache.cassandra.db.Keyspace.<init>(Keyspace.java:272)
> at org.apache.cassandra.db.Keyspace.open(Keyspace.java:114)
> at org.apache.cassandra.db.Keyspace.open(Keyspace.java:92)
> at
> org.apache.cassandra.db.SystemKeyspace.checkHealth(SystemKeyspace.java:536)
> at
> org.apache.cassandra.service.CassandraDaemon.setup(CassandraDaemon.java:261)
> at
> org.apache.cassandra.service.CassandraDaemon.activate(CassandraDaemon.java:496)
> at
> org.apache.cassandra.service.CassandraDaemon.main(CassandraDaemon.java:585)
> {code}
> I looked at source code of Cassandra and see:
> http://grepcode.com/file/repo1.maven.org/maven2/org.apache.cassandra/cassandra-all/2.0.10/org/apache/cassandra/db/RowIndexEntry.java
> 119 int entries = in.readInt();
> 120 List<IndexHelper.IndexInfo> columnsIndex = new
> ArrayList<IndexHelper.IndexInfo>(entries);
> It seems that value entries is invalid (negative) and it tries too allocate
> an array with huge initial capacity and hits OOM. I have deleted saved_cache
> directory and was able to start node correctly. We should expect that it may
> happen in real world. Cassandra should be able to skip incorrect cached data
> and run.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
