Fast and Simplified Streaming,Ad-Hoc and Batch Analytics with

FiloDB and Spark Streaming and Evan Chan Helena Edelson

March 2016

Evan Chan

Distinguished Engineer,

User and contributor to Spark since 0.9, Cassandra since0.6Co-creator and maintainer of

Tuplejump@evanfchanhttp://velvia.github.io

Spark Job Server

Helena EdelsonVP of Product Engineering, Committer: Kafka Connect Cassandra, Spark CassandraConnectorContributor to Akka (2 new features in Akka Cluster), SpringIntegrationSpeaker @ Spark Summit, Kafka Summit, Strata, QCon, ScalaDays, Scala World, Philly ETE

Tuplejump

@helenaedelsonhttp://github.com/helena

Tuplejump is a big data technology leader providing solutions and

development partnership.Tuplejump

Open Source: on GitHubTuplejump - Subject of today's talk

- Kafka-Cassandra Source andSink

- The �rst Spark Cassandra integration - Lucene indexer for Cassandra - HDFS for Cassandra

FiloDBKafka Connect Cassandra

CalliopeStargateSnackFS

Tuplejump Consulting & Development

Tuplejump BlenderBuilds uni�ed datasets for fast querying, streaming and batch sources, ML and Analytics

TopicsModern streaming architectures and batch/ad-hocarchitecturesPrecise and scalable streaming ingestion using Kafka, Akka,Spark Streaming, Cassandra, and FiloDBHow a uni�ed streaming + batch stack can lower your TCOWhat FiloDB is and how it enables fast analytics withcompetitive storage costData Warehousing with Spark, Cassandra, and FiloDBTime series / event data / geospatial examplesMachine learning using Spark MLLib + Cassandra/FiloDB

The Problem DomainBuild scalable, adaptable, self-healing, distributed dataprocessing systems for

Massive amounts of dataDisparate sources and schemasDiffering data structuresComplex analytics and learning tasksTo run as large-scale clustereddata�ows24-7 UptimeGlobally clustered deploymentsNo data loss

Delivering MeaningDeliver meaning in sec/sub-sec latencyBillions of events per secondLow latency real time stream processingHigher latency batch processingAggregation of global data from thestream

While We Monitor, Predict & ProactivelyHandle

Massive event spikes & bursty traf�cFast producers / slow consumersNetwork partitioning & out of syncsystemsDC downNot DDOS'ing ourselves from fast streamsNo data loss when auto-scaling down

And stay within ourAWS

OpenStackRackspace...

budget

Use CaseI need fast access to historical data on the �y for predictive

modeling with real time data from the stream

Only, It's Not A Stream It's A FloodNet�ix

100 billion events per day1-2 million events per second atpeak

Linkedin

500 billion write events per day2.5 trillion read events per day4.5 million events per second at peak withKafkaPetabytes of streaming data

Con�dential

700 TB global ingested data (pre-�ltered)

Lambda ArchitectureA data-processing architecture designed to handle massive quantitiesof data by taking advantage of both batch and stream processingmethods.

Lambda Architecture

( )https://www.mapr.com/developercentral/lambda-architecture

Challenge Assumptions"Ingest an immutable sequence of records is captured and fed into

a batch processing systemand a stream processing system

in parallel"

Lambda ArchitectureDual Analytics Systems

Many moving parts: KV store, real time, Batch technologiesRunning similar code and reconciling queries in dualsystems

Overly Complicated

PipelinesOps: e.g. performance tuning & monitoring, upgrades...Analytics logic changes: dev/testing/deploys of changes tocode bases, clusters

Ultimately High TCOAnd...

Technical Debt

Are Batch and Streaming SystemsFundamentally Different?

No.

A Unified Streaming ArchitectureEverything On The Streaming Platform

Scala / SparkStreamingMesosAkkaCassandraKafka

SNACK (SMACK) Stack

Apache Spark Streaming

One runtime for streaming and batch processingJoin streaming and static data setsNo code duplicationEasy Kafka stream integrationEasy to reconcile queries against multiplesourcesEasy integration of KV durable storage

High Throughput Distributed MessagingDecouples Data PipelinesHandles Massive Data LoadSupport Massive Number of ConsumersDistribution & partitioning across clusternodesAutomatic recovery from broker failures

Apache Cassandra

Horizontally scalableMulti-Region / Multi-DatacenterAlways On - Survive regional outagesExtremely fast writes: - perfect for ingestion of real time /machine dataVery �exible data modelling (lists, sets, custom data types)Easy to operateBest of breed storage technology, huge communityBUT: Simple queries onlyOLTP-oriented/center

High performance concurrency framework for Scala andJavaFault ToleranceAsynchronous messaging and data processingParallelizationLocation TransparencyLocal / Remote RoutingAkka: Cluster / Persistence / Streams

EnablesStreaming and Batch In One System

Streaming ML and Analytics for Predictions In The Stream

Show me the code

Immutable Raw Data From Kafka StreamReplay / reprocessing: for fault tolerance, logic changes..class KafkaStreamingActor(ssc: StreamingContext, settings: Settings) extends AggregationActor val stream = KafkaUtils.createDirectStream(...).map(RawWeatherData(_)) stream .foreachRDD(_.toDF.write.format("filodb.spark")) .option(rawDataKeyspace, rawDataTable) /* Pre-Aggregate data in the stream for fast querying and aggregation later. */ stream.map(hour => (hour.wsid, hour.year, hour.month, hour.day, hour.oneHourPrecip) ).saveToCassandra(CassandraKeyspace, CassandraTableDailyPrecip) }

Reading Data Back From CassandraCompute isolation in Akka Actor

class TemperatureActor(sc: SparkContext, settings: Settings) extends AggregationActor import settings._ import akka.pattern.pipe def receive: Actor.Receive = { case e: GetMonthlyHiLowTemperature => highLow(e, sender) } def highLow(e: GetMonthlyHiLowTemperature, requester: ActorRef): Unit = sc.cassandraTable[DailyTemperature](timeseriesKeyspace, dailyTempAggregTable) .where("wsid = ? AND year = ? AND month = ?", e.wsid, e.year, e.month) .collectAsync() .map(MonthlyTemperature(_, e.wsid, e.year, e.month)) pipeTo requester }

Spark Streaming, MLLibKafka, Cassandra

val ssc = new StreamingContext(sparkConf, Seconds(5) val testData = ssc.cassandraTable[String](keyspace,table) .map(LabeledPoint.parse) val trainingStream = KafkaUtils.createDirectStream[_,_,_,_](..) .map(transformFunc) .map(LabeledPoint.parse) trainingStream.saveToCassandra("ml_training_keyspace", "raw_training_data") val model = new StreamingLinearRegressionWithSGD() .setInitialWeights(Vectors.dense(weights)) .trainOn(trainingStream) model .predictOnValues(testData.map(lp => (lp.label, lp.features))) .saveToCassandra("ml_predictions_keyspace", "predictions")

Tuplejump OSS Roadmap gearpump-external-

�lodbKafka Connect FiloDB

Gearpump

What's Missing? One Pipeline For Fast +Big Data

Using Cassandra for Batch Analytics /Event Storage / ML?

Storage ef�ciency and scan speeds for reading large volumesof data (for complex analytics, ML) become importantconcernsRegular Cassandra CQL tables are not very good at eitherstorage ef�ciency or scan speedsA different, analytics-optimized solution is needed...

All hard work leads to pro�t, but mere talk leadsto poverty.- Proverbs 14:23

Introducing FiloDBA distributed, versioned, columnar analytics database.

Built for Streaming.

 

github.com/tuplejump/FiloDB

Fast Analytics StorageScan speeds competitive with Apache Parquet

Up to 200x faster scan speeds than with Cassandra 2.xFlexible �ltering along two dimensions

Much more ef�cient and �exible partition key �lteringEf�cient columnar storage, up to 27x more ef�cient thanCassandra 2.x

Comparing Storage Costs and Query Speeds

https://www.oreilly.com/ideas/apache-cassandra-for-analytics-a-performance-and-storage-analysis

Robust Distributed StorageApache Cassandra as the rock-solid storage engine. Scale outwith no SPOF. Cross-datacenter replication. Proven storage anddatabase technology.

 

Cassandra-Like Data ModelColumn A Column B

Partitionkey 1

Segment1

Segment2

Segment1

Segment2

Partitionkey 2

Segment1

Segment2

Segment1

Segment2

partition keys - distributes data around a cluster, and allowsfor �ne grained and �exible �lteringsegment keys - do range scans within a partition, e.g. by timesliceprimary key based ingestion and updates

Flexible FilteringUnlike Cassandra, FiloDB offers very �exible and ef�cient�ltering on partition keys. Partial key matches, fast IN queries onany part of the partition key.

No need to write multiple tables to work around answering differentqueries.

Spark SQL Queries!CREATE TABLE gdelt USING filodb.spark OPTIONS (dataset "gdelt"); SELECT Actor1Name, Actor2Name, AvgTone FROM gdelt ORDER BY AvgTone DESC LIMIT 15 INSERT INTO gdelt SELECT * FROM NewMonthData;

Read to and write from Spark DataframesAppend/merge to FiloDB table from SparkStreamingUse Tableau or any other JDBC tool

What's in the name?

Rich sweet layers of distributed, versioned database goodness

SNACK (SMACK) stack for all yourAnalytics

Regular Cassandra tables for highly concurrent, aggregate /key-value lookups (dashboards)FiloDB + C* + Spark for ef�cient long term event storage

Ad hoc / SQL / BIData source for MLLib / building modelsData storage for classi�ed / predicted / scored data

Being Productionized as we speak...One enterprise with many TB of �nancial and reporting data ismoving their data warehouse to FiloDB + Cassandra + SparkAnother startup uses FiloDB as event storage, feeds the eventsinto Spark MLlib, scores incoming data, then stores the resultsback in FiloDB for low-latency use cases

From their CTO: “I see close to MemSQL / Vertica or evenbetter” “More cost effective than Redshift”

Data Warehousing with FiloDB

ScenariosBI Reporting, concurrency + seconds latencyAd-hoc queriesNeeding to do JOINs with fact tables + dimensiontables

Slowly changing dim tables / hard to denormalizeNeed to work with legacy BI tools

Real-world DW Architecture Stack

Ef�cient columnar storage + �ltering = low latency BI

Modeling Fact Tables for FiloDBSingle partition queries are really fast and take up only onethread

Given the following two partition key columns:entity_number, year_monthWHERE entity_number = '0453' ANDyear_month = '2014 December'

Exact match for partition key is pushed down as onepartition

Consider the partition key carefully

Cassandra often requires multiple tablesWhat about the queries that do not translate to one partition?Cassandra has many restrictions on partition key �ltering (as of2.x).

Table 1: partition key = (entity_number, year_month)Can push down: WHERE entity_number = NN ANDyear_month IN ('2014 Jan', '2014 Feb') aswell as equals

Table 2: partition key = (year_month, entity_number)Can push down: WHERE year_month = YYMM ANDentity_number IN (123, 456) as well as equals

IN clause must be the last column to be pushed down. Two tablesare needed just for ef�cient IN queries on either entity_numberor year_month.

FiloDB Flexible Partition Filters = WINWith ONE table, FiloDB offers FAST, arbitrary partition key�ltering. All of the below are pushed down:

WHERE year_month IN ('2014 Jan', '2014 Feb')(all entities)WHERE entity_number = 146 (all year months)Any combo of =, IN

Space savings: 27 * 2 = 54x

Multi-Table JOINs with just Cassandra

Sub-second Multi-Table JOINs with FiloDB

Sub-second Multi-Table JOINs with FiloDBFour tables, all of them single-partition queriesTwo tables were switched from regular Cassandra tables toFiloDB tables. 40-60 columns each, ~60k items in partition.Scan times went down from 5-6 seconds to < 250ms

For more details, please see this .Planet Cassandra blog post

Scalable Time-Series / Event Storage withFiloDB

Designed for StreamingNew rows appended via Spark Streaming or KafkaWrites are idempotent - easy exactly once ingestionConverted to columnar chunks on ingest and stored inC*FiloDB keeps your data sorted as it is being ingested

Spark Streaming -> FiloDB val ratingsStream = KafkaUtils.createDirectStream[String, String, StringDecoder, StringDecoder](ssc, kafkaParams, topics) ratingsStream.foreachRDD { (message: RDD[(String, String)], batchTime: Time) => { val df = message.map(_._2.split(",")).map(rating => Rating(rating(0).trim.toInt, rating( toDF("fromuserid", "touserid", "rating") // add the batch time to the DataFrame val dfWithBatchTime = df.withColumn("batch_time", org.apache.spark.sql.functions.lit(batchTime.milliseconds)) // save the DataFrame to FiloDB dfWithBatchTime.write.format("filodb.spark") .option("dataset", "ratings") .save() } }

One-line change to write to FiloDB vs Cassandra

Modeling example: NYC Taxi DatasetThe public contains telemetry (pickup, dropofflocations, times) info on millions of taxi rides in NYC.

NYC Taxi Dataset

Medallion pre�x 1/1 - 1/6 1/7 - 1/12AA records records

AB records recordsPartition key - :stringPrefix medallion 2 - hashmultiple drivers trips into ~300 partitionsSegment key - :timeslice pickup_datetime 6dRow key - hack_license, pickup_datetime

Allows for easy �ltering by individual drivers, and slicing by time.

DEMO TIMENew York City Taxi Data Demo (Spark Notebook)

To follow along:https://github.com/tuplejump/FiloDB/blob/master/doc/FiloDB_Taxi_Geo_demo.snb

Fast, Updatable In-MemoryColumnar Storage

Unlike RDDs and DataFrames, FiloDB can ingest new data, andstill be fastUnlike RDDs, FiloDB can �lter in multiple ways, no need forentire table scanFAIR scheduler + sub-second latencies => web speed queries

700 Queries Per Second in Apache Spark!Even for datasets with 15 million rows!Using FiloDB's InMemoryColumnStore, single host / MBP,5GB RAMSQL to DataFrame caching

For more details, see .this blog post

Machine Learning with Spark, Cassandra,and FiloDB

Building a static model of NYC Taxi TripsPredict time to get to destination based on pickup point, timeof day, other varsNeed to read all data (full table scan)

Dynamic models are better than staticmodels

Everything changes!Continuously re�ne model based on recent streaming data +historical data + existing model

val ssc = new StreamingContext(sparkConf, Seconds(5)) val dataStream = KafkaUtils.createDirectStream[..](..) .map(transformFunc) .map(LabeledPoint.parse) dataStream.foreachRDD(_.toDF.write.format("filodb.spark") .option("dataset", "training").save()) if (trainNow) { var model = new StreamingLinearRegressionWithSGD() .setInitialWeights(Vectors.dense(weights)) .trainOn(dataStream.join(historicalEvents)) } model.predictOnValues(dataStream.map(lp => (lp.label, lp.features))) .insertIntoFilo("predictions")

The FiloDB Advantage for MLAble to update dynamic models based on massive data�ow/updates

Integrate historical and recent events to build modelsMore data -> better models!Can store scored raw data / predictions back in FiloDB

for fast user queries

FiloDB - RoadmapYour input is appreciated!

Productionization and automated stress testingKafka input API / connector (without needing Spark)In-memory caching for signi�cant query speedupTrue columnar querying and execution, using latematerialization and vectorization techniques. GPU/SIMD.Projections. Often-repeated queries can be sped upsigni�cantly with projections.

Thanks For Attending!@helenaedelson@evanfchan@tuplejump

EXTRA SLIDES

What are my storage needs?Non-persistent / in-memory: concurrentviewersShort term: latest trendsLonger term: raw event and aggregate storageML Models, predictions, scored data

Spark RDDsImmutable, cache in memory and/or ondiskSpark Streaming: UpdateStateByKeyIndexedRDD - can update bits of dataSnapshotting for recovery

Using Cassandra for Short Term Storage1020s 1010s 1000s

Bus A Speed, GPS

Bus B

Bus CPrimary key = (Bus UUID, timestamp)Easy queries: location and speed of single bus for a range oftimeCan also query most recent location + speed of all buses(slower)

FiloDB - How?

Multiple ways to Accelerate QueriesColumnar projection - read fewer columns, saves I/OPartition key �ltering - read less dataSort key / PK �ltering - read from subset of keys

Possible because FiloDB keeps data sortedVersioning - write to multiple versions, read from the one youchoose

Cassandra CQL vs Columnar LayoutCassandra stores CQL tables row-major, each row spans multiplecells:

PartitionKey 01:�rst 01:last 01:age 02:�rst 02:last 02:ageSales Bob Jones 34 Susan O'Connor 40

Engineering Dilbert P ? Dogbert Dog 1 

Columnar layouts are column-major:

PartitionKey �rst last ageSales Bob, Susan Jones,

O'Connor34,40

Engineering Dilbert,Dogbert

P, Dog ?, 1

FiloDB Cassandra SchemaCREATE TABLE filodb.gdelt_chunks ( partition text, version int, columnname text, segmentid blob, chunkid int, data blob, PRIMARY KEY ((partition, version), columnname, segmentid, chunkid) ) WITH CLUSTERING ORDER BY (columnname ASC, segmentid ASC, chunkid ASC)

FiloDB Architecture

ColumnStore API - currently Cassandra and InMemory, you canimplement other backends - ElasticSearch? etc.