[jira] [Comment Edited] (CASSANDRA-6405) When making heavy use of counters, neighbor nodes occasionally enter spiral of constant memory consumpion

["Jason Harvey (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/CASSANDRA-6405?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13861360#comment-13861360 ] 

Jason Harvey edited comment on CASSANDRA-6405 at 1/3/14 8:59 AM:
-----------------------------------------------------------------

[~slebresne] When the issue is occurring, we have no pending ReplicateOnWrite threads. All pending threads are either reads or writes.

When the crazy CounterColumn instance counts are reached, the number of reads/writes occurring on the table are drastically reduced they're all moving extremely slowly. 

If the CounterColumn instance count was legitimate, wouldn't we expect a huge number of reads/writes to be occurring, rather than a small few? Even during peak hours, we don't do more than 150 reads / 10 writes a second per cassandra node. When this issue occurs, that drops down to 2-3 reads and writes a second.

Additionally, we allow up to 128 read threads concurrently. Most of the counter column rows have around 1k columns, with the 95th percentile having 18k columns. Even if every single read thread was dedicated to reading our largest countercolumn row (which they're not), that accounts for a maximum of ~2-3m counter columns being concurrently accessed.


was (Author: alienth):
[~slebresne] When the issue is occurring, we have no pending ReplicateOnWrite threads. All pending threads are either reads or writes.

When the issue is occurring and the crazy CounterColumn instance counts are reached, the number of reads/writes occurring on the table are drastically reduced they're all moving extremely slowly. 

If the CounterColumn instance count was legitimate, wouldn't we expect a huge number of reads/writes to be occurring, rather than a small few? Even during peak hours, we don't do more than 150 reads / 10 writes a second per cassandra node. When this issue occurs, that drops down to 2-3 reads and writes a second.

Additionally, we allow up to 128 read threads concurrently. Most of the counter column rows have around 1k columns, with the 95th percentile having 18k columns. Even if every single read thread was dedicated to reading our largest countercolumn row (which they're not), that accounts for a maximum of ~2-3m counter columns being concurrently accessed.

> When making heavy use of counters, neighbor nodes occasionally enter spiral of constant memory consumpion
> ---------------------------------------------------------------------------------------------------------
>
>                 Key: CASSANDRA-6405
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-6405
>             Project: Cassandra
>          Issue Type: Bug
>         Environment: RF of 3, 15 nodes.
> Sun Java 7 (also occurred in OpenJDK 6, and Sun Java 6).
> Xmx of 8G.
> No row cache.
>            Reporter: Jason Harvey
>         Attachments: threaddump.txt
>
>
> We're randomly running into an interesting issue on our ring. When making use of counters, we'll occasionally have 3 nodes (always neighbors) suddenly start immediately filling up memory, CMSing, fill up again, repeat. This pattern goes on for 5-20 minutes. Nearly all requests to the nodes time out during this period. Restarting one, two, or all three of the nodes does not resolve the spiral; after a restart the three nodes immediately start hogging up memory again and CMSing constantly.
> When the issue resolves itself, all 3 nodes immediately get better. Sometimes it reoccurs in bursts, where it will be trashed for 20 minutes, fine for 5, trashed for 20, and repeat that cycle a few times.
> There are no unusual logs provided by cassandra during this period of time, other than recording of the constant dropped read requests and the constant CMS runs. I have analyzed the log files prior to multiple distinct instances of this issue and have found no preceding events which are associated with this issue.
> I have verified that our apps are not performing any unusual number or type of requests during this time.
> This behaviour occurred on 1.0.12, 1.1.7, and now on 1.2.11.
> The way I've narrowed this down to counters is a bit naive. It started happening when we started making use of counter columns, went away after we rolled back use of counter columns. I've repeated this attempted rollout on each version now, and it consistently rears its head every time. I should note this incident does _seem_ to happen more rarely on 1.2.11 compared to the previous versions.
> This incident has been consistent across multiple different types of hardware, as well as major kernel version changes (2.6 all the way to 3.2). The OS is operating normally during the event.
> I managed to get an hprof dump when the issue was happening in the wild. Something notable in the class instance counts as reported by jhat. Here are the top 5 counts for this one node:
> {code}
> 5967846 instances of class org.apache.cassandra.db.CounterColumn 
> 1247525 instances of class com.googlecode.concurrentlinkedhashmap.ConcurrentLinkedHashMap$WeightedValue 
> 1247310 instances of class org.apache.cassandra.cache.KeyCacheKey 
> 1246648 instances of class com.googlecode.concurrentlinkedhashmap.ConcurrentLinkedHashMap$Node 
> 1237526 instances of class org.apache.cassandra.db.RowIndexEntry 
> {code}
> Is it normal or expected for CounterColumn to have that number of instances?
> The data model for how we use counters is as follows: between 50-20000 counter columns per key. We currently have around 3 million keys total, but this issue also replicated when we only had a few thousand keys total. Average column count is around 1k, and 90th is 18k. New columns are added regularly, and columns are incremented regularly. No column or key deletions occur. We probably have 1-5k "hot" keys at any given time, spread across the entire ring. R:W ratio is typically around 50:1. This is the only CF we're using counters on, at this time. CF details are as follows:
> {code}
>     ColumnFamily: CommentTree
>       Key Validation Class: org.apache.cassandra.db.marshal.AsciiType
>       Default column value validator: org.apache.cassandra.db.marshal.CounterColumnType
>       Cells sorted by: org.apache.cassandra.db.marshal.CompositeType(org.apache.cassandra.db.marshal.LongType,org.apache.cassandra.db.marshal.LongType,org.apache.cassandra.db.marshal.LongType)
>       GC grace seconds: 864000
>       Compaction min/max thresholds: 4/32
>       Read repair chance: 0.01
>       DC Local Read repair chance: 0.0
>       Populate IO Cache on flush: false
>       Replicate on write: true
>       Caching: KEYS_ONLY
>       Bloom Filter FP chance: default
>       Built indexes: []
>       Compaction Strategy: org.apache.cassandra.db.compaction.LeveledCompactionStrategy
>       Compaction Strategy Options:
>         sstable_size_in_mb: 160
>                 Column Family: CommentTree
>                 SSTable count: 30
>                 SSTables in each level: [1, 10, 19, 0, 0, 0, 0, 0, 0]
>                 Space used (live): 4656930594
>                 Space used (total): 4677221791
>                 SSTable Compression Ratio: 0.0
>                 Number of Keys (estimate): 679680
>                 Memtable Columns Count: 8289
>                 Memtable Data Size: 2639908
>                 Memtable Switch Count: 5769
>                 Read Count: 185479324
>                 Read Latency: 1.786 ms.
>                 Write Count: 5377562
>                 Write Latency: 0.026 ms.
>                 Pending Tasks: 0
>                 Bloom Filter False Positives: 2914204
>                 Bloom Filter False Ratio: 0.56403
>                 Bloom Filter Space Used: 523952
>                 Compacted row minimum size: 30
>                 Compacted row maximum size: 4866323
>                 Compacted row mean size: 7742
>                 Average live cells per slice (last five minutes): 39.0
>                 Average tombstones per slice (last five minutes): 0.0
> {code}
> Please let me know if I can provide any further information. I can provide the hprof if desired, however it is 3GB so I'll need to provide it outside of JIRA.



--
This message was sent by Atlassian JIRA
(v6.1.5#6160)