OpenTSDB + Bigtable Integrating time series database with Google - - PowerPoint PPT Presentation

OpenTSDB + Bigtable

Integrating time series database with Google Cloud Bigtable

Danil Zburivsky, Big Data Practice Lead - @zburivsky Christos Soulios, Big Data Architect - @c_soulios

Pythian specializes in design, implementation, and management of systems that directly contribute to revenue and business success. History 19 years in business Growing at 30+% per year 400+ employees 300+ customers worldwide HQ Ottawa, Canada - global reach Technology agnostic = trusted advisor Deep expertise: Oracle, Oracle Apps, MySQL, AWS, SQL Server, Cassandra/DataStax, Azure, PostgreSQL, Cloudera, MapR, Hortonworks etc. Google Premier Partner Status (as of end Aug) 5 Certified Developers (soon to be 12) Dedicated Google Technical Champion Launch partner for: Kubernetes, Dataflow, Cloud SQL, Dataproc Integrated OpenTSDB with Bigtable DW Explorers Program Partner Upcoming BigQuery & Cloud ML Launch Partner

• (time, metric, value)
• OS and apps metrics
• Industrial equipment
• Web traffic

Time series data

• Volume can be explosive
• Data arrival and access

patterns are different

Storing time series data is a challenge

• Volume can be explosive
• Data arrival and access

patterns are different

Storing time series data is a challenge

• NoSQL
• Data model and storage optimized for time series
• Separate query language

Better alternatives — specialized stores

• Open source
• Uses HBase as a data store
• Data model optimized for time series
• REST API

OpenTSDB

<metric_uid><timestamp><tagk1><tagv1>[...<tagkN><tagvN>] <col_t+1>[...<col_t+N>]

OpenTSDB Architecture

Server Server Server Server TSD TSD HBase

TSD RPC HBase RPC Web UI Scripts/Alerting HTTP TSD RPC

• HBase requires a full Hadoop setup

(3xZK, 2xNN, 3xDN, 2xHMaster, 3xHRegion)

• HBase tuning is a job for the brave

(HFiles, WAL, MemStore, BucketCache, BlockCache)

HBase can be too much

HBase can be too much

But all I wanted was a time series database

Google Cloud Bigtable

• Highly Scalable NoSQL database
• Low latency, high throughput
• Powers most Google products
• Available as a Google Cloud Service

Migrate HBase apps to Cloud Bigtable

• The Bigtable client is API compatible with HBase client
• Only replace hbase-client.jar with bigtable-hbase.jar
• No code changes required!

Migrate OpenTSDB to Cloud Bigtable

• OpenTSDB does not use standard hbase-client.jar
• OpenTSDB is based on AsyncHBase library

AsyncHBase library

• Open source HBase client library
• Multi-threaded

Multiple threads use the same instance

• Fully asynchronous, non-blocking
• Implements the low level HBase RPCs

Detour: Asynchronous programming

Detour: Why asynchronous?

• Efficient thread usage
• Less threads = less memory
• CPU scheduler friendly
• Extremely high concurrency

AsyncHBase library

http://www.tsunanet.net/~tsuna/asynchbase/benchmark/viz.html

AsyncHBase library

“AsyncHBase client differs significantly from HBase's

• client. Switching to it is not easy as it requires to rewrite all

the code that was interacting with any HBase API”

AsyncHBase documentation

AsyncBigtable library

• Complete rewrite of AsyncHBase API
• Uses standard hbase-client for Bigtable access
• Compatible with the bigtable-hbase API

AsyncBigtable challenges

• OpenTSDB jar dependencies
• AsyncBigtable is not async!
• BufferedMutator + Threadpool to emulate async

AsyncBigtable library

AsyncBigtable library

• Merged upstream OpenTSDB v2.3.0
• http://opentsdb.net/docs/build/html/user_guide/backend

s/bigtable.html

• https://github.com/OpenTSDB/asyncbigtable

Future work

• Native Bigtable API
• Fully asynchronous
• Improve performance
• Add more unit tests

Questions?

https://github.com/opentsdb/asyncbigtable