Managing a Cassandra Cluster: Lessons Learned from 3M+ Node-Hours of Experience 19 April 2016
Agenda • About Instaclustr & presenters • Foundation practices for a happy cluster • The most common Cassandra issues and how to avoid them • Important monitoring and health-check procedures • Q & A
About Instaclustr • Cassandra, Spark and Zeppelin Managed Service on AWS, Azure and SoftLayer • 500+ nodes • >80 clusters from 3 nodes to 50+ • >3M node-hours of C* experience
• Cassandra and Spark Consulting & Support
• Ben Slater – Chief Product Officer
• Aleks Lubiejewski – VP Consulting & Security
(prev VP Support)
What do I need to get right at the beginning to have a happy cluster?
Pick the right Cassandra version • Most stable -> Cassandra 2.1 (or, better yet, DSE) • Want the latest features and can live with “cutting edge” or have a few months until production -> Cassandra 3.x • Odd numbers (eg 3.5) should be more stable as these are defect-fix releases
• Cassandra 2.2: same end of life as Cassandra 2.1, no DSE version (yet?). Only recommended if you really want the new features but don’t want to jump to 3.x.
Appropriate Hardware Configuration • Solid State Disks • Lots of memory • Can use EBS on AWS (and equivalent on other platforms) but needs very careful attention to configuration
• For the cloud, we prefer more, smaller nodes: • In AWS, m4.xlarge with 800G or 1.6TB are our standard building blocks • Smaller proportionate impact of failure of a node • Reasonable time to replace a failed node or add new ones
Estimating Costs in the Cloud Cost Description Driver Instances Cost of the base compute instances (eg m4.xl). Number and size of nodes in the cluster. EBS Volume Cost of attached EBS volumes (where applicable) Size of the EBS volume (eg 400GB) Network – Public IP Loading/retrieving data via public IP Only applicable if accessing via Public IP: dependant on number In/Out of Cassandra read/writes in a month and transaction size. Network – Interzone Cross-availability zone communication within the Transaction volume and size, consistency factor used for reads In/Out cluster Network – Loading/retrieving data via a peered VPC Only applicable if accessing via Peered VPC: dependant on VPC In/Out number of Cassandra read/writes in a month and transaction size. S3 Storage S3 space for storing backups Volume of data, length of backup retention, deduplication of backup files/data S3 Operations S3 calls for storing backups Number of sstables (volume of data + compaction strategy), backup strategy S3 Data Transfer Out S3 retrieval data transfer cost Only applicable if you need to copy data from S3 to a region other than US East to restore a backup. • EBS and network costs can exceed instance cost in some circumstances
Cassandra–Specific Load Testing • Short load tests from the application side can be misleading. • Need to consider: • Is there enough data on disk to overflow the operating system file cache? • Does your data reflect production distributions, particularly for primary key fields? • Are you performing deletes/updates to test for impact of tombstones (virtual deletes)? • If you are using cassandra-stress, do you understand the available options and their impacts?
NetworkTopologyStrategy, RF=3,CL=Quorum • NetworkTopologyStrategy • Most “future proof” strategy • Allows use of multi-dc • Can be very useful for cluster operations such as upgrades and splitting out tables
• Replication Factor = 3, Consistency Level = Quorom • Provides expected behaviour for most people: • Strong consistency • Availability through failure of a replica • Other settings are valid for many uses cases but need careful consideration of the impacts
Most Common Issues
What are the most common causes when things go wrong in production?
Data Modelling Issues • Partition Keys • Cassandra primary keys consist of a partition key and a clustering key. • Eg PRIMARY KEY ((c1, c2), c3, c4) • Partition keys determine how data is distributed around the nodes • 1 node per partition, many partitions per node • Need to ensure there are are a large number of partitions with a reasonable number of rows per partitions • Small number or very uneven partitions defeat the basic concepts of Cassandra scaling • Very large partitions can causes issues with garbage collection and excess disk usage
Data Modelling Issues (2) • Tombstones • Tombstones are entries created to mark the fact that a record has been deleted (updates to primary key will also cause tombstones) • By default, tombstones are retained for 10 days before being removed by compactions • High ratios of tombstones to live data can cause significant performance issues • Secondary Indexes • Secondary Indexes are useful for a limited set of cases • only index low (but not too low) cardinality columns; • don’t index columns that are frequently updated or deleted • Poor use of secondary indexes can result in poor read performance and defeat scalability
Other Issues • Write Overload • Cluster may be able to initial handle a write workload but then fail to keep up as compactions kick in • The effects can go beyond slow latency and cause crashes • Garbage Collection (GC) • Long garbage collection pauses can often cause issues • Typically a symptom of partitions being too big or general overload of cluster • Tuning GC settings and heap allocations can help but need to address root cause also
Running out of capacity • Cassandra scales infinitely but processing capacity is used when adding new nodes to a cluster. • Therefore, you need to add capacity well before existing capacity is exhausted. • This applies to both disk and processor/memory.
Fundamental Monitoring & Health Check Procedures
How do I get advanced warning if my cluster is going to hit issues?
Monitoring – Basic Metrics (OS) • Disk usage • less than 70% under normal running is a good starting guide • this allows for bursts of usage by compactions and repairs • extreme cases may required 50% free space • Levelled compaction strategy and data spread across multiple column families can allow higher disk usage • CPU Usage • again, 70% utilization is a reasonable target • Keep a look out for high iowait – indicates storage bottleneck
Monitoring – Basic Metrics (C*) • Read/Write Latency • closely tied to user experience for most use cases • monitor for significant changes • be aware that read latency can very greatly depending on number of rows returned • distinguish between changes impacting a specific column family (likely data modelling issues) and changes impacting a specific node (hardware or capacity issues) • Pending Compactions • increase numbers of pending compactions indicates that a node is not keeping up with workload
Cassandra Logs • Regularly inspecting Cassandra logs for warnings and errors is important for picking up issues. The issues you will find include: • large batch warnings
Apr 18 08:00:27 ip-10-224-111-138.ec2.internal docker: [Native-Transport-Requests:22756] WARN org.apache.cassandra.cql3.statements.BatchStatement Batch of prepared statements for [ks.col_family] is of size
59000, exceeding specified threshold of 5120 by 53880. • compacting large partitions warnings ip-10-224 -169-153.eu-west-1.compute.internal docker: WARN o.a.c.io.sstable.SSTableWriter Compacting large partition ks/col_family:c7d65814-1a58-4675-ad54-d6c92e10d1d7 (405357404 bytes)
• reading excess tombstones warnings Mar 29 11 :55:26 ip-172-16-151-148.ec2.internal docker: WARN o.a.c.db.filter.SliceQueryFilter Read 3563 live and 3520 tombstone cells in ks.col_family for key: 4364012 (see tombstone_warn_threshold). 5000 columns were requested , slices=[2016-03-28 11¥:55Z:!-] • plenty more!
cfstats and cfhistograms • The Cassandra nodetool tool has many commands that help diagnose issues. • nodetool cfstats and cfhistograms are two of the most important • These tools can help you see: • Large and uneven partitions • Excess tombstones • Too many sstables per read • read and write latency by keyspace and column family • many more
Cassandra is an incredibly reliable and scalable technology
you design and build correctly from the start and follow basic management procedures.