Any way to control/limit off-heap memory?

2017-03-04 Thread Thakrar, Jayesh 
I have a situation where the off-heap memory is bloating the jvm process 
memory, making it a candidate to be killed by the oom_killer.
My server has 256 GB RAM and Cassandra heap memory of 16 GB

Below is the output of "nodetool info" and nodetool compactionstats for a 
culprit table which causes bloom filter bloat.
Ofcourse one option is to turnoff bloom filter, but I need to look into 
application access pattern, etc.


xss =  -ea -Dorg.xerial.snappy.tempdir=/home/vchadoop/var/tmp 
-javaagent:/home/vchadoop/apps/apache-cassandra-2.2.5//lib/jamm-0.3.0.jar 
-XX:+UseThreadPriorities -XX:ThreadPriorityPolicy=42 -Xms16G -Xmx16G -Xmn4800M 
-XX:+HeapDumpOnOutOfMemoryError -Xss256k
ID : 2b9b4252-0760-49c1-8d14-544be0183271
Gossip active  : true
Thrift active  : false
Native Transport active: true
Load   : 953.19 GB
Generation No  : 1488641545
Uptime (seconds)   : 15706
Heap Memory (MB)   : 7692.93 / 16309.00
Off Heap Memory (MB)   : 175115.07
Data Center: ord
Rack   : rack3
Exceptions : 0
Key Cache  : entries 0, size 0 bytes, capacity 0 bytes, 0 hits, 0 
requests, NaN recent hit rate, 14400 save period in seconds
Row Cache  : entries 0, size 0 bytes, capacity 0 bytes, 0 hits, 0 
requests, NaN recent hit rate, 0 save period in seconds
Counter Cache  : entries 0, size 0 bytes, capacity 50 MB, 0 hits, 0 
requests, NaN recent hit rate, 7200 save period in seconds
Token  : (invoke with -T/--tokens to see all 256 tokens)


Table: logs_by_user
SSTable count: 622
SSTables in each level: [174/4, 447/10, 0, 0, 0, 0, 0, 0, 0]
Space used (live): 313156769247
Space used (total): 313156769247
Space used by snapshots (total): 0
Off heap memory used (total): 180354511884
SSTable Compression Ratio: 0.25016314078395613
Number of keys (estimate): 147261312
Memtable cell count: 44796
Memtable data size: 57578717
Memtable off heap memory used: 0
Memtable switch count: 21
Local read count: 0
Local read latency: NaN ms
Local write count: 1148687
Local write latency: 0.123 ms
Pending flushes: 0
Bloom filter false positives: 0
Bloom filter false ratio: 0.0
Bloom filter space used: 180269125192
Bloom filter off heap memory used: 180269120216
Index summary off heap memory used: 24335340
Compression metadata off heap memory used: 61056328
Compacted partition minimum bytes: 150
Compacted partition maximum bytes: 668489532
Compacted partition mean bytes: 3539
Average live cells per slice (last five minutes): NaN
Maximum live cells per slice (last five minutes): 0
Average tombstones per slice (last five minutes): NaN
Maximum tombstones per slice (last five minutes): 0


From: Conversant

Re: OOM on Apache Cassandra on 30 Plus node at the same time

2017-03-04 Thread Thakrar, Jayesh 
If possible, I would suggest running that command on a periodic basis (cron or 
whatever).
Also, you can run it on a single server and iterate through all the nodes in 
the cluster/DC.
Would also recommend running "nodetool compactionstats

And looked at your concern about high value for hinted handoff.
That's good (in a way), it ensures that updates are not lost.
Its possible because your DB was constantly being updated and the servers that 
come up started accumulating for the servers that were still down.
Furthermore, that may have been the situation also as the servers were going 
down.
Hence high hinted handoff is just a sign of pending updates that need to be 
applied, which is not uncommon if you had servers falling down/restarting like 
dominos and updates still coming in.

From: Shravan C 
Date: Saturday, March 4, 2017 at 11:15 AM
To: Conversant , Joaquin Casares 
, "user@cassandra.apache.org" 

Subject: Re: OOM on Apache Cassandra on 30 Plus node at the same time


I was looking at nodetool info across all nodes. Consistently JVM heap used is 
~ 12GB and off heap is ~ 4-5GB.


From: Thakrar, Jayesh 
Sent: Saturday, March 4, 2017 9:23:01 AM
To: Shravan C; Joaquin Casares; user@cassandra.apache.org
Subject: Re: OOM on Apache Cassandra on 30 Plus node at the same time

LCS does not rule out frequent updates - it just says that there will be more 
frequent compaction, which can potentially increase compaction activity (which 
again can be throttled as needed).
But STCS will guarantee OOM when you have large datasets.
Did you have a look at the offheap + onheap size of our jvm using "nodetool 
-info" ?


From: Shravan C 
Date: Friday, March 3, 2017 at 11:11 PM
To: Joaquin Casares , "user@cassandra.apache.org" 

Subject: Re: OOM on Apache Cassandra on 30 Plus node at the same time


We run C* at 32 GB and all servers have 96GB RAM. We use STCS . LCS is not an 
option for us as we have frequent updates.


Thanks,
Shravan

From: Thakrar, Jayesh 
Sent: Friday, March 3, 2017 3:47:27 PM
To: Joaquin Casares; user@cassandra.apache.org
Subject: Re: OOM on Apache Cassandra on 30 Plus node at the same time


Had been fighting a similar battle, but am now over the hump for most part.



Get info on the server config (e.g. memory, cpu, free memory (free -g), etc)

Run "nodetool info" on the nodes to get heap and off-heap sizes

Run "nodetool tablestats" or "nodetool tablestats ." on the 
key large tables

Essentially the purpose is to see if you really had a true OOM or was your 
machine running out of memory.



Cassandra can use offheap memory very well - so "nodetool info" will give you 
both heap and offheap.



Also, what is the compaction strategy of your tables?



Personally, I have found STCS to be awful at large scale - when you have 
sstables that are 100+ GB in size.

See 
https://issues.apache.org/jira/browse/CASSANDRA-10821?focusedCommentId=15389451=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-15389451



LCS seems better and should be the default (my opinion) unless you want DTCS



A good description of all three compactions is here - 
http://docs.scylladb.com/kb/compaction/
Documentation
docs.scylladb.com
Scylla is a Cassandra-compatible NoSQL data store that can handle 1 million 
transactions per second on a single server.








From: Joaquin Casares 
Date: Friday, March 3, 2017 at 11:34 AM
To: 
Subject: Re: OOM on Apache Cassandra on 30 Plus node at the same time



Hello Shravan,



Typically asynchronous requests are recommended over batch statements since 
batch statements will cause more work on the coordinator node while individual 
requests, when using a TokenAwarePolicy, will hit a specific coordinator, 
perform a local disk seek, and return the requested information.



The only times that using batch statements are ideal is if writing to the same 
partition key, even if it's across multiple tables when using the same hashing 
algorithm (like murmur3).



Could you provide a bit of insight into what the batch statement was trying to 
accomplish and how many child statements were bundled up within that batch?



Cheers,



Joaquin


Joaquin Casares

Consultant

Austin, TX



Apache Cassandra Consulting

http://www.thelastpickle.com
The Last Pickle • Apache Cassandra Consulting & 
Services
www.thelastpickle.com
Apache Cassandra Consulting & Services. Our wealth of experience with Apache 
Cassandra will ensure success at all stages of a your project lifecycle.




On Fri, Mar 3, 2017 at 11:18 AM, Shravan Ch

Re: OOM on Apache Cassandra on 30 Plus node at the same time

2017-03-04 Thread Priyanka 


Sent from my iPhone

> On Mar 3, 2017, at 12:18 PM, Shravan Ch  wrote:
> 
> Hello,
> 
> More than 30 plus Cassandra servers in the primary DC went down OOM exception 
> below. What puzzles me is the scale at which it happened (at the same 
> minute). I will share some more details below. 
> 
> System Log: http://pastebin.com/iPeYrWVR
> GC Log: http://pastebin.com/CzNNGs0r
> 
> During the OOM I saw lot of WARNings like the below (these were there for 
> quite sometime may be weeks)
> WARN  [SharedPool-Worker-81] 2017-03-01 19:55:41,209 BatchStatement.java:252 
> - Batch of prepared statements for [keyspace.table] is of size 225455, 
> exceeding specified threshold of 65536 by 159919.
> 
> Environment:
> We are using ApacheCassandra-2.1.9 on Multi DC cluster. Primary DC (more C* 
> nodes on SSD and apps run here)  and secondary DC (geographically remote and 
> more like a DR to primary) on SAS drives. 
> Cassandra config:
> 
> Java 1.8.0_65
> Garbage Collector: G1GC
> memtable_allocation_type: offheap_objects
> 
> Post this OOM I am seeing huge hints pile up on majority of the nodes and the 
> pending hints keep going up. I have increased HintedHandoff CoreThreads to 6 
> but that did not help (I admit that I tried this on one node to try).
> 
> nodetool compactionstats -H
> pending tasks: 3
> compaction typekeyspace  table   
> completed  totalunit   progress
> Compaction  system  hints 
> 28.5 GB   92.38 GB   bytes 30.85%
> 
> 
> Appreciate your inputs here. 
> 
> Thanks,
> Shravan

Re: OOM on Apache Cassandra on 30 Plus node at the same time

2017-03-04 Thread Shravan C 
I was looking at nodetool info across all nodes. Consistently JVM heap used is 
~ 12GB and off heap is ~ 4-5GB.


From: Thakrar, Jayesh 
Sent: Saturday, March 4, 2017 9:23:01 AM
To: Shravan C; Joaquin Casares; user@cassandra.apache.org
Subject: Re: OOM on Apache Cassandra on 30 Plus node at the same time

LCS does not rule out frequent updates - it just says that there will be more 
frequent compaction, which can potentially increase compaction activity (which 
again can be throttled as needed).
But STCS will guarantee OOM when you have large datasets.
Did you have a look at the offheap + onheap size of our jvm using "nodetool 
-info" ?


From: Shravan C 
Date: Friday, March 3, 2017 at 11:11 PM
To: Joaquin Casares , "user@cassandra.apache.org" 

Subject: Re: OOM on Apache Cassandra on 30 Plus node at the same time


We run C* at 32 GB and all servers have 96GB RAM. We use STCS . LCS is not an 
option for us as we have frequent updates.


Thanks,
Shravan

From: Thakrar, Jayesh 
Sent: Friday, March 3, 2017 3:47:27 PM
To: Joaquin Casares; user@cassandra.apache.org
Subject: Re: OOM on Apache Cassandra on 30 Plus node at the same time


Had been fighting a similar battle, but am now over the hump for most part.



Get info on the server config (e.g. memory, cpu, free memory (free -g), etc)

Run "nodetool info" on the nodes to get heap and off-heap sizes

Run "nodetool tablestats" or "nodetool tablestats ." on the 
key large tables

Essentially the purpose is to see if you really had a true OOM or was your 
machine running out of memory.



Cassandra can use offheap memory very well - so "nodetool info" will give you 
both heap and offheap.



Also, what is the compaction strategy of your tables?



Personally, I have found STCS to be awful at large scale - when you have 
sstables that are 100+ GB in size.

See 
https://issues.apache.org/jira/browse/CASSANDRA-10821?focusedCommentId=15389451=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-15389451



LCS seems better and should be the default (my opinion) unless you want DTCS



A good description of all three compactions is here - 
http://docs.scylladb.com/kb/compaction/
Documentation
docs.scylladb.com
Scylla is a Cassandra-compatible NoSQL data store that can handle 1 million 
transactions per second on a single server.








From: Joaquin Casares 
Date: Friday, March 3, 2017 at 11:34 AM
To: 
Subject: Re: OOM on Apache Cassandra on 30 Plus node at the same time



Hello Shravan,



Typically asynchronous requests are recommended over batch statements since 
batch statements will cause more work on the coordinator node while individual 
requests, when using a TokenAwarePolicy, will hit a specific coordinator, 
perform a local disk seek, and return the requested information.



The only times that using batch statements are ideal is if writing to the same 
partition key, even if it's across multiple tables when using the same hashing 
algorithm (like murmur3).



Could you provide a bit of insight into what the batch statement was trying to 
accomplish and how many child statements were bundled up within that batch?



Cheers,



Joaquin


Joaquin Casares

Consultant

Austin, TX



Apache Cassandra Consulting

http://www.thelastpickle.com
The Last Pickle • Apache Cassandra Consulting & 
Services
www.thelastpickle.com
Apache Cassandra Consulting & Services. Our wealth of experience with Apache 
Cassandra will ensure success at all stages of a your project lifecycle.




On Fri, Mar 3, 2017 at 11:18 AM, Shravan Ch 
> wrote:

Hello,

More than 30 plus Cassandra servers in the primary DC went down OOM exception 
below. What puzzles me is the scale at which it happened (at the same minute). 
I will share some more details below.

System Log: http://pastebin.com/iPeYrWVR

GC Log: http://pastebin.com/CzNNGs0r

During the OOM I saw lot of WARNings like the below (these were there for quite 
sometime may be weeks)
WARN  [SharedPool-Worker-81] 2017-03-01 19:55:41,209 BatchStatement.java:252 - 
Batch of prepared statements for [keyspace.table] is of size 225455, exceeding 
specified threshold of 65536 by 159919.

Environment:
We are using ApacheCassandra-2.1.9 on Multi DC cluster. Primary DC (more C* 
nodes on SSD and apps run here)  and secondary DC (geographically remote and 
more like a DR to primary) on SAS drives.
Cassandra config:

Java 1.8.0_65
Garbage Collector: G1GC
memtable_allocation_type: offheap_objects

Post this OOM I am seeing huge hints pile up on majority of the nodes and the 
pending hints keep going up. I have increased HintedHandoff CoreThreads to 6 
but

Re: OOM on Apache Cassandra on 30 Plus node at the same time

2017-03-04 Thread Edward Capriolo 
On Saturday, March 4, 2017, Thakrar, Jayesh 
wrote:

> LCS does not rule out frequent updates - it just says that there will be
> more frequent compaction, which can potentially increase compaction
> activity (which again can be throttled as needed).
>
> But STCS will guarantee OOM when you have large datasets.
>
> Did you have a look at the offheap + onheap size of our jvm using
> "nodetool -info" ?
>
>
>
>
>
> *From: *Shravan C  >
> *Date: *Friday, March 3, 2017 at 11:11 PM
> *To: *Joaquin Casares  >, "
> user@cassandra.apache.org
> " <
> user@cassandra.apache.org
> >
> *Subject: *Re: OOM on Apache Cassandra on 30 Plus node at the same time
>
>
>
> We run C* at 32 GB and all servers have 96GB RAM. We use STCS . LCS is not
> an option for us as we have frequent updates.
>
>
>
> Thanks,
>
> Shravan
> --
>
> *From:* Thakrar, Jayesh  >
> *Sent:* Friday, March 3, 2017 3:47:27 PM
> *To:* Joaquin Casares; user@cassandra.apache.org
> 
> *Subject:* Re: OOM on Apache Cassandra on 30 Plus node at the same time
>
>
>
> Had been fighting a similar battle, but am now over the hump for most part.
>
>
>
> Get info on the server config (e.g. memory, cpu, free memory (free -g),
> etc)
>
> Run "nodetool info" on the nodes to get heap and off-heap sizes
>
> Run "nodetool tablestats" or "nodetool tablestats ."
> on the key large tables
>
> Essentially the purpose is to see if you really had a true OOM or was your
> machine running out of memory.
>
>
>
> Cassandra can use offheap memory very well - so "nodetool info" will give
> you both heap and offheap.
>
>
>
> Also, what is the compaction strategy of your tables?
>
>
>
> Personally, I have found STCS to be awful at large scale - when you have
> sstables that are 100+ GB in size.
>
> See https://issues.apache.org/jira/browse/CASSANDRA-10821?
> focusedCommentId=15389451=com.atlassian.jira.
> plugin.system.issuetabpanels:comment-tabpanel#comment-15389451
>
>
>
> LCS seems better and should be the default (my opinion) unless you want
> DTCS
>
>
>
> A good description of all three compactions is here -
> http://docs.scylladb.com/kb/compaction/
>
> Documentation 
>
> docs.scylladb.com
>
> Scylla is a Cassandra-compatible NoSQL data store that can handle 1
> million transactions per second on a single server.
>
>
>
>
>
>
>
> *From: *Joaquin Casares  >
> *Date: *Friday, March 3, 2017 at 11:34 AM
> *To: * >
> *Subject: *Re: OOM on Apache Cassandra on 30 Plus node at the same time
>
>
>
> Hello Shravan,
>
>
>
> Typically asynchronous requests are recommended over batch statements
> since batch statements will cause more work on the coordinator node while
> individual requests, when using a TokenAwarePolicy, will hit a specific
> coordinator, perform a local disk seek, and return the requested
> information.
>
>
>
> The only times that using batch statements are ideal is if writing to the
> same partition key, even if it's across multiple tables when using the same
> hashing algorithm (like murmur3).
>
>
>
> Could you provide a bit of insight into what the batch statement was
> trying to accomplish and how many child statements were bundled up within
> that batch?
>
>
>
> Cheers,
>
>
>
> Joaquin
>
>
> Joaquin Casares
>
> Consultant
>
> Austin, TX
>
>
>
> Apache Cassandra Consulting
>
> http://www.thelastpickle.com
>
> The Last Pickle • Apache Cassandra Consulting & Services
> 
>
> www.thelastpickle.com
>
> Apache Cassandra Consulting & Services. Our wealth of experience with
> Apache Cassandra will ensure success at all stages of a your project
> lifecycle.
>
>
>
> On Fri, Mar 3, 2017 at 11:18 AM, Shravan Ch  > wrote:
>
> Hello,
>
> More than 30 plus Cassandra servers in the primary DC went down OOM
> exception below. What puzzles me is the scale at which it happened (at the
> same minute). I will share some more details below.
>
> System Log: http://pastebin.com/iPeYrWVR
>
> GC Log: http://pastebin.com/CzNNGs0r
>
> During the OOM I saw lot of WARNings like the below (these were there for
> quite sometime may be weeks)
> *WARN  [SharedPool-Worker-81] 2017-03-01 19:55:41,209
> BatchStatement.java:252 - Batch of prepared statements for [keyspace.table]
> is of size 225455, exceeding

Re: OOM on Apache Cassandra on 30 Plus node at the same time

2017-03-04 Thread Thakrar, Jayesh 
LCS does not rule out frequent updates - it just says that there will be more 
frequent compaction, which can potentially increase compaction activity (which 
again can be throttled as needed).
But STCS will guarantee OOM when you have large datasets.
Did you have a look at the offheap + onheap size of our jvm using "nodetool 
-info" ?


From: Shravan C 
Date: Friday, March 3, 2017 at 11:11 PM
To: Joaquin Casares , "user@cassandra.apache.org" 

Subject: Re: OOM on Apache Cassandra on 30 Plus node at the same time


We run C* at 32 GB and all servers have 96GB RAM. We use STCS . LCS is not an 
option for us as we have frequent updates.


Thanks,
Shravan

From: Thakrar, Jayesh 
Sent: Friday, March 3, 2017 3:47:27 PM
To: Joaquin Casares; user@cassandra.apache.org
Subject: Re: OOM on Apache Cassandra on 30 Plus node at the same time


Had been fighting a similar battle, but am now over the hump for most part.



Get info on the server config (e.g. memory, cpu, free memory (free -g), etc)

Run "nodetool info" on the nodes to get heap and off-heap sizes

Run "nodetool tablestats" or "nodetool tablestats ." on the 
key large tables

Essentially the purpose is to see if you really had a true OOM or was your 
machine running out of memory.



Cassandra can use offheap memory very well - so "nodetool info" will give you 
both heap and offheap.



Also, what is the compaction strategy of your tables?



Personally, I have found STCS to be awful at large scale - when you have 
sstables that are 100+ GB in size.

See 
https://issues.apache.org/jira/browse/CASSANDRA-10821?focusedCommentId=15389451=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-15389451



LCS seems better and should be the default (my opinion) unless you want DTCS



A good description of all three compactions is here - 
http://docs.scylladb.com/kb/compaction/
Documentation
docs.scylladb.com
Scylla is a Cassandra-compatible NoSQL data store that can handle 1 million 
transactions per second on a single server.








From: Joaquin Casares 
Date: Friday, March 3, 2017 at 11:34 AM
To: 
Subject: Re: OOM on Apache Cassandra on 30 Plus node at the same time



Hello Shravan,



Typically asynchronous requests are recommended over batch statements since 
batch statements will cause more work on the coordinator node while individual 
requests, when using a TokenAwarePolicy, will hit a specific coordinator, 
perform a local disk seek, and return the requested information.



The only times that using batch statements are ideal is if writing to the same 
partition key, even if it's across multiple tables when using the same hashing 
algorithm (like murmur3).



Could you provide a bit of insight into what the batch statement was trying to 
accomplish and how many child statements were bundled up within that batch?



Cheers,



Joaquin


Joaquin Casares

Consultant

Austin, TX



Apache Cassandra Consulting

http://www.thelastpickle.com
The Last Pickle • Apache Cassandra Consulting & 
Services
www.thelastpickle.com
Apache Cassandra Consulting & Services. Our wealth of experience with Apache 
Cassandra will ensure success at all stages of a your project lifecycle.




On Fri, Mar 3, 2017 at 11:18 AM, Shravan Ch 
> wrote:

Hello,

More than 30 plus Cassandra servers in the primary DC went down OOM exception 
below. What puzzles me is the scale at which it happened (at the same minute). 
I will share some more details below.

System Log: http://pastebin.com/iPeYrWVR

GC Log: http://pastebin.com/CzNNGs0r

During the OOM I saw lot of WARNings like the below (these were there for quite 
sometime may be weeks)
WARN  [SharedPool-Worker-81] 2017-03-01 19:55:41,209 BatchStatement.java:252 - 
Batch of prepared statements for [keyspace.table] is of size 225455, exceeding 
specified threshold of 65536 by 159919.

Environment:
We are using ApacheCassandra-2.1.9 on Multi DC cluster. Primary DC (more C* 
nodes on SSD and apps run here)  and secondary DC (geographically remote and 
more like a DR to primary) on SAS drives.
Cassandra config:

Java 1.8.0_65
Garbage Collector: G1GC
memtable_allocation_type: offheap_objects

Post this OOM I am seeing huge hints pile up on majority of the nodes and the 
pending hints keep going up. I have increased HintedHandoff CoreThreads to 6 
but that did not help (I admit that I tried this on one node to try).

nodetool compactionstats -H
pending tasks: 3
compaction typekeyspace  table   completed  
totalunit   progress
Compaction  system  hints 28.5 
GB   92.38 GB   bytes 30.85%


Appreciate your inputs here.