Cassandra at Pollfish
2 likes852 views
The document outlines the technology and architecture used by Pollfish, particularly focusing on their use of Apache Cassandra and Spark for handling surveys targeted at mobile users. It details the cluster setup, best practices for performance optimization, and integration with analytics frameworks. Additionally, it highlights the advantages and disadvantages of using Cassandra and DSE, along with tips for managing and scaling the cluster effectively.
Cassandra at Pollfish
- 1. Stavros Kontopoulos Software Engineer, MSc Cassandra at Pollfish.com
- 3. What we do at Pollfish? • Target mobile users with surveys through our android/ ios sdk which is installed via thousands of mobile apps. Developers benefit from completed surveys, companies also may run a survey campaign in real time. Now analytics/ML pipeline… Why Apache Cassandra? • Store time series about system events, user activities, survey results and much more.. • Amazing write throughput, take advantage of idempotent writes with proper resolution. • Decent read speed throughput and low latency. • Integrates with spark to implement our analytics, insights pipeline. • A new business model. A flashback… why I am here…
- 4. Our Tech Team- Disciplinary oriented • Front End Development - UI Design • Back-end, data engineer(s) • Data Scientist • DevOps
- 5. Pollfish High Level Architecture Mobile Users (~600K active per day) APP SERVER 1 APP SERVER N Other systems (PostGres, Geo, Redis…) DSE Cassandra/Spark Cluster Survey Customers
- 6. Why Datastax Start up program and support when you go to production Many tools for development and maintenance: Pig, hive, shark, CFS, OpsCenter. Clear setup and support for real time data storage and analytics in the same cluster. Can be extended for other workloads like Search (Solr). Supports multiple DCs easily for other purposes staging, backup etc. Product version we used: DSE 4.5.1 (spark 0.9.1), 4.5.3, 4.6.0 (Spark 1.1.0)
- 7. Our Cassandra Cluster- Setup 2 DataCenters (Cassandra Analytics): • 1 for real-time data storage, read/write path. • 1 for analytics nodes (Spark and Hadoop enabled), read/write path ETL, machine learning. • Use the DSE setup with DSESimpleSnitch for mixed workloads multi-DC clusters. 3 Nodes per DC (details next), planning for more. Data is written at Cassandra DC and replicated to Analytics DC for all keyspaces needed. 2 seed nodes, one per DC.
- 8. Our Cassandra Cluster- Setup
- 9. Our Cassandra Cluster- Setup
- 10. Our Cassandra Cluster- Infrastructure Details Cassandra DC: • Node 0,1,2: Standard_A4 (8 cores, 14 GB memory), 15DISKs in RAID 0 Analytics DC: • Node 0,1,2: Standard_D13 (8 cores, 56 GB memory), 15DISKs in RAID 0 A bit about disks (ideally JBOD s but we can only have network disks)… FS type: Ext4 Analytics disks per node: Azure Os disk (/dev/sda1 , mp: /): 29GB Azure VM tmp disk (/dev/sdb1, mp: /mnt/resource): 400GB - for spark temp files… data_file_directories: /mnt/dsedata/lib/cassandra/data (2,9T) commitlog_directory: /var/lib/cassandra/commitlog (197GB disk)
- 11. A few tips towards production… Follow as a start: http://www.datastax.com/documentation/cassandra/2.0/cassandra/install/installRecommendSettings.html http://www.datastax.com/wp-content/uploads/2014/04/WP-DataStax-Enterprise-Best-Practices.pdf • NTP is a must for all servers, Cassandra needs that for node data synchronization and when you do analytics • Proper limits: cassandra - memlock unlimited cassandra - nofile 100000 cassandra - nproc 32768 cassandra - as unlimited • Optimum blockdev --setra settings for RAID (OK): All value must be set to 128KB for RA. • No swap
- 12. A few Cassandra optimization tips For Cassandra DC: • concurrent_reads: 64 • concurrent_writes: 32 • Adjust according to the load and node IOPS/throughput.. Heap size adjustments: Datastax does a good job to automatically handle this… In general for production 8GB for heap is ok… to avoid gc pauses… found in bigger jvms. We use 4GB in smaller Cassandra nodes and 8GB in Analytics. • In case you have a load with short lived objects adjust accordingly the parNew size. Your tools: jconsole,jvisualvm,hprof, jstack, enable GC reports at logging level… JVM_OPTS=”$JVM_OPTS -XX:+PrintGCDetails” JVM_OPTS=”$JVM_OPTS -XX:+PrintGCDateStamps” JVM_OPTS=”$JVM_OPTS -XX:+PrintHeapAtGC“ …
- 13. Managing nodes When you have a running car how you change the tire? Adding a new clean node to the cluster fast : • Add the following line to /etc/dse/cassandra/cassandra-env.sh : • JVM_OPTS=”$JVM_OPTS -Dcassandra.join_ring=false“ • Start the service • Revert change and execute nodetool join • After joining finishes execute a node rebalance from OpsCenter. • Fallback: remove data and commit log dirs. And start again. Stopping a node: nodetool drain -h host name sudo service dse stop
- 14. Integration with Spark- Use cases Commmon ETL case: Mobile User profile (ML): HDFS Compatible (CFS) Cassandra Raw Data Spark/Cassandra Spark/Cassandra HDFS Compatible (CFS) Cassandra User profile data Customers Amazon S3
- 15. Integration with Spark- Our job framework • We have built an analytics framework to run spark jobs on top of Cassandra. Executes jobs remotely through maven. • SparkContext is created on a non-dse node. Needed some tricks to make it work, now we move to a more dse official approach with spark submit. Note: We started with 0.9.1 DSE version where no spark submit was available. • The framework provides an API to run jobs in production and use common functionality at the Data science side for code re-use.
- 16. Datastax Cassandra Opscenter • Facilitates: monitoring, health check and daily operations at the cluster level. • We use it for: node rebalance, repairs,snapshotting, latency throughput checks along with low level tools like iotop, ioping etc.
- 17. Datastax Cassandra Opscenter (Pending cluster operations)
- 18. Datastax Cassandra Opscenter (Keyspace Details)
- 19. Datastax Cassandra Opscenter (Last 10 days)
- 20. The Good and the Bad with Cassandra and DSE good • Easy to start with, enough documentation. • Reliable performance so far. • Easily scalable, easy to add/remove nodes. • Support. bad • Bugs at the Cassandra level and the Spark level. Need active follow up of the lists for Spark, Cassandra and DSE tools. Upgrade maybe the only solution and not a piece of case. • Gets tricky to optimize depending on your load specifically when you don’t have time to measure everything upfront and you are in production already.
- 21. The Good and the Bad with Cassandra and DSE (Some issue examples) Apply backpressure gently when overloaded with writes, https://issues.apache.org/jira/browse/CASSANDRA-7937 Reproduce by writing large volume of data with large column values eg. 300K, crawling scenario... Spark Cassandra Connector issues (eg. java driver not closing connection with cluster)...
- 22. As a developer some cool stuff I mess with Currently I am extending the framework I have built to deliver spark jobs on top of Cassandra. • IT with Cassandra embedded connector. • Cassandra Schema Design. • ETL use cases. Organizing dev env: Optimizing development environment with vagraant, automatic installation - avoid upgrade headaches, setup and IT level testing. Introduced CI with Jenkins, Artifactory etc…
- 23. thank you