I would like to know about hardware limitations in cluster planning (in TBs) specific to my use case. I have read few threads and documents related to it but some content seem to be over 5 years old. Thought of giving it a shot again:
Use case: Building a time-series cassandra cluster where there is from time-to-time bulk loading from data sources which are in Gigabytes. However, the end-user will majorly be focused in reading the data from the cluster. Quite rarely will be some update or delete on the rows
I have an initial hardware configuration with me to setup Cassandra cluster:
2*12 Cores
128 GB RAM
HDD SAS 3.27 TB
This is the initial plan that I come up with:
When I now speculate over the setup, and after reading the post:
should I further divide my nodes with lesser RAM, vCPUs and HDD?
If yes, what would be the good fit wrt my case?
Hi the official Spark documentation state:
While Spark can perform a lot of its computation in memory, it still
uses local disks to store data that doesn’t fit in RAM, as well as to
preserve intermediate output between stages. We recommend having 4-8
disks per node, configured without RAID (just as separate mount
points). In Linux, mount the disks with the noatime option to reduce
unnecessary writes. In Spark, configure the spark.local.dir variable
to be a comma-separated list of the local disks. If you are running
HDFS, it’s fine to use the same disks as HDFS.
I wonder what is the purpose of 4-8 per node
Is it for parallel write ? I am not sure to understand the reason why as it is not explained.
I have no clue for this: "If you are running HDFS, it’s fine to use
the same disks as HDFS".
Any idea what is meant here...
Purpose of usage 4-8 RAID disks to mirror partitions adding redundancy to prevent data lost in case of fault on hardware level. In case of HDFS the redundancy that RAID provides is not needed, since HDFS handles it by replication between nodes.
Reference
I deployed a Cassandra 2.2 ring composed by 4 nodes in the cloud with 8 vCPU and 8GB of ram. I am running some tests now with cassandra-stress and YCSB tools to test its performance. I am mainly interested in read requests with a small amount of write requests (95%/5%).
Running the experiments, I noticed that even setting a high number of threads (or clients) the CPU (and disk) does not saturate, but still always around the 60% of utilisation.
I am trying to figure out where is the bottleneck in my system. From the hardware point of view it seems all ok to me.
dstat
I also looked into the Cassandra configuration file to see if there are some tuning parameters to increase the system throughput. I increase the value of concurrent_read/write parameter, but it doesn't increase the performance.
The log file also does not contain any warning.
What it could be that is limiting my system?
Thanks
You might want to consider running cassandra-stress from outside the cluster and on multiple instances as described in
Usage of the Cassandra tool cassandra-stress
I started learning Apache Cassandra. In the conf/cassandra.yaml I noticed the commitlog setting's comment as following:
commit log. when running on magnetic HDD, this should be a
separate spindle than the data directories.
If not set, the default directory is $CASSANDRA_HOME/data/commitlog.
Does that mean I should store the commitlog in different HDD than the data?
If yes, what's the reason behind this? And what will happen if I don't comply.
Thanks.
That is a recommendation from the days of spinning-disk. Due to its log-based storage engine, Cassandra is very dependent on disk I/O. So it was recommended to have your commit log and data directories on separate disks to avoid a potential bottleneck (latency) due to heavy disk activity.
If you are using solid state drives (SSDs, and with Cassandra you really should be) then you don't need to worry about this.
NOTE: This is also why using a NAS or SAN with Cassandra is considered to be an anti-pattern.
I'm running Datastax Enterprise in a cluster consisting of 3 nodes. They are all running under the same hardware: 2 Core Intel Xeon 2.2 Ghz, 7 GB RAM, 4 TB Raid-0
This should be enough for running a cluster with a light load, storing less than 1 GB of data.
Most of the time, everything is just fine but it appears that sometimes the running tasks related to the Repair Service in OpsCenter sometimes get stuck; this causes an instability in that node and an increase in load.
However, if the node is restarted, the stuck tasks don't show up and the load is at normal levels again.
Because of the fact that we don't have much data in our cluster we're using the min_repair_time parameter defined in opscenterd.conf to delay the repair service so that it doesn't complete too often.
It really seems a little bit weird that the tasks that says that are marked as "Complete" and are showing a progress of 100% don't go away, and yes, we've waited hours for them to go away but they won't; the only way that we've found to solve this is to restart the nodes.
Edit:
Here's the output from nodetool compactionstats
Edit 2:
I'm running under Datastax Enterprise v. 4.6.0 with Cassandra v. 2.0.11.83
Edit 3:
This is output from dstat on a node that behaving normally
This is output from dstat on a node with stucked compaction
Edit 4:
Output from iostat on node with stucked compaction, see the high "iowait"
azure storage
Azure divides disk resources among storage accounts under an individual user account. There can be many storage accounts in an individual user account.
For the purposes of running DSE [or cassandra], it is important to note that a single storage account should not should not be shared between more than two nodes if DSE [or cassandra] is configured like the examples in the scripts in this document. This document configures each node to have 16 disks. Each disk has a limit of 500 IOPS. This yields 8000 IOPS when configured in RAID-0. So, two nodes will hit 16,000 IOPS and three would exceed the limit.
See details here
So, this has been an issue that have been under investigation for a long time now and we've found a solution, however, we aren't sure what the underlaying problem that were causing the issues were but we got a clue even tho that, nothing can be confirmed.
Basically what we did was setting up a RAID-0 also known as Striping consisting of four disks, each at 1 TB of size. We should have seen somewhere 4x one disks IOPS when using the Stripe, but we didn't, so something was clearly wrong with the setup of the RAID.
We used multiple utilities to confirm that the CPU were waiting for the IO to respond most of the time when we said to ourselves that the node was "stucked". Clearly something with the IO and most probably our RAID-setup was causing this. We tried a few differences within MDADM-settings etc, but didn't manage to solve the problems using the RAID-setup.
We started investigating Azure Premium Storage (which still is in preview). This enables attaching disks to VMs whose underlaying physical storage actually are SSDs. So we said, well, SSDs => more IOPS, so let us give this a try. We did not setup any RAID using the SSDs. We are only using one single SSD-disk per VM.
We've been running the Cluster for almost 3 days now and we've stress tested it a lot but haven't been able to reproduce the issues.
I guess we didn't came down to the real cause but the conclusion is that some of the following must have been the underlaying cause for our problems.
Too slow disks (writes > IOPS)
RAID was setup incorrectly which caused the disks to function non-normally
These two problems go hand-in-hand and most likely is that we basically just was setting up the disks in the wrong way. However, SSDs = more power to the people, so we will definitely continue using SSDs.
If someone experience the same problems that we had on Azure with RAID-0 on large disks, don't hesitate to add to here.
Part of the problem you have is that you do not have a lot of memory on those systems and it is likely that even with only 1GB of data per node, your nodes are experiencing GC pressure. Check in the system.log for errors and warnings as this will provide clues as to what is happening on your cluster.
The rollups_60 table in the OpsCenter schema contains the lowest (minute level) granularity time series data for all your Cassandra, OS, and DSE metrics. These metrics are collected regardless of whether you have built charts for them in your dashboard so that you can pick up historical views when needed. It may be that this table is outgrowing your small hardware.
You can try tuning OpsCenter to avoid this kind of issues. Here are some options for configuration in your opscenterd.conf file:
Adding keyspaces (for example the opsc keyspace) to your ignored_keyspaces setting
You can also decrease the TTL on this table by tuning the 1min_ttlsetting
Sources:
Opscenter Config DataStax docs
Metrics Config DataStax Docs