Spark After Dark - LA Apache Spark Users Group - Feb 2015
29 likes5,069 views
The document presents insights into various technologies and paradigms related to Spark, including Spark Streaming, Machine Learning, Graph Processing, and the Lambda Architecture. It details the functionalities, use cases, and enhancements of Spark's capabilities and APIs, alongside methods for deploying Spark in cluster environments. Key topics include execution optimizations, fault tolerance, and the integration of the Spark ecosystem with other big data technologies.
![Message Delivery Guarantees
• Exactly once [1]!
– No loss!
– No redeliver!
– Perfect delivery!
– Incurs higher latency for transactional semantics!
– Spark default per batch using DStream lineage!
– Degrades to less guarantees depending on source!
• At least once [1..n]!
– No loss!
– Possible redeliver!
• At most once [0,1]!
– Possible loss!
– No redeliver!
– *Best configuration if some data loss is acceptable!
• Ordered!
– Per partition: Kafka, Kinesis!
– Global across all partitions: Hard to scale!](https://image.slidesharecdn.com/sparkafterdark-150221200547-conversion-gate01/85/Spark-After-Dark-LA-Apache-Spark-Users-Group-Feb-2015-94-320.jpg)
Spark After Dark - LA Apache Spark Users Group - Feb 2015
- 1. after! Dark! ! Spark Streaming, Machine Learning, Graph Processing, Lambda Architecture, Probabilistic Data Structs/Algos, Approximations! ! Chris Fregly!
- 2. Who am I? Former Netflix’er,! NetflixOSS Committer! (netflix.github.io) ! ! Spark Contributor! (github.com/apache/spark)! ! Author! (sparkinaction.com)! (effectivespark.com)! ! ! ! ! Field Engineering! (databricks.com)!
- 4. Not-so-Quick Poll • Execution Engines! – Spark, Hadoop, Hive, Pig! • File Formats! – Parquet, Avro, RCFile, ORCFile! • Streaming! – Spark Streaming, Storm, Flume, Kafka, Kinesis ! • AWS Big Data! – EMR, DynamoDB, Redshift, Kinesis! • Lambda Architecture! – Combined Batch and Real-time Processing! • Machine Learning! – Recommendations, Clustering, Classification! • Graph Processing! – PageRank, Nearest Neighbor, Triangle Count! • Approximations! – Bloom Filters, HyperLogLog, Count-min Sketch!
- 5. Agenda • Spark After Dark Overview! • Spark Overview! • Use Cases! • API and Libraries! • Machine Learning! • Graph Processing! • Execution Model! • Fault Tolerance! • Cluster Deployment! • Monitoring! • Scaling and Tuning! • Lambda Architecture! • Probabilistic Data Structs/Algos! • Approximations! DEMOS!!! DEMOS!!! DEMOS!!! DEMOS!!! DEMOS!!! DEMOS!!! DEMOS!!! DEMOS!!! DEMOS!!! DEMOS!!! DEMOS!!! DEMOS!!! DEMOS!!! DEMOS!!! DEMOS!!!
- 6. Spark After Dark (SAD) Overview! sparkafterdark.com “Better Dating through Algorithms”!
- 7. Spark After Dark (SAD) ~ Tinder* Swipe Left (Reject) Swipe Right (Accept) *not affiliated with Tinder in any way
- 8. Spark After Dark Sample Dataset
- 10. Matches: Spark SQL, BlinkDB • Traditional rating aggregations! – Most-active users (SUM)! – Most-rated users (SUM)! – Top-rated users (AVG)! – Spark SQL! • Approximate online rating aggregations! – Error-bound, time-bound! – When exact results aren’t needed! – Near-real-time trend analysis! – BlinkDB!
- 11. Matches: MLlib, GraphX • Find similar based on! interests, religion, etc! – MLlib: K-Means Cluster, ! DIMSUM Similarity, ! DecisionTree & ! RandomForest Classifier! ! • Recommendations! – Fill in missing ratings! – MLlib: Alternating Least ! Squares (ALS)!
- 12. Matches: MLlib + GraphX • Profile Text Analysis! – MLlib: TF/IDF,! N-gram, LDA Topics! - GraphX: TextRank! • Top Influencers! – Most desirable people! – GraphX: PageRank! ! • People You May Know! – More intimate matches! – GraphX: Shortest Path,! Connected Components!
- 13. Matches: Spark Streaming • Capture real-time, online, incremental trends!! • Spark Streaming + MLlib + GraphX! – Clustering: StreamingKMeans.scala! – Logistics Regression: StreamingLogR.scala! – Linear Regression: StreamingLinR.scala! – Naïve-Bayes Classification: TODO! – DecisionTree/RandomForest: TODO! – PageRank: TODO! • Spark Streaming + Spark SQL! – Lambda Architecture: real-time + batch
- 14. Spark Overview
- 15. What is Spark? • Based on 2007 Microsoft Dryad paper! • Written in Scala! • Supports Java, Python, SQL, and R! • Data fits in memory when possible, but not required! • Improved efficiency over MapReduce! – 100x in-memory, 2-10x on-disk! • Compatible with Hadoop! – File formats, SerDes, UDFs, Hive, and Pig!
- 16. Timeline of Spark Evolution
- 17. Spark and Berkeley AMP Lab! Berkeley Data Analytics Stack (BDAS)!
- 18. Advantages of Unified Libraries • Advancements in higher-level libraries are pushed down into core and vice-versa! • Spark Core! – Highly-optimized, low overhead, network-saturating shuffle! • Spark Streaming! – GC, memory management, and cleanup improvements ! • Spark GraphX! – IndexedRDD for random access within a partition vs scanning entire partition! • MLlib ! – Statistics (Correlations, sampling, heuristics, etc)!
- 19. Hadoop Support! • Shark! – Hive on Spark! – Used Hive execution engine – didn’t allow for optimizations! – *Deprecated in favor of Spark SQL + Thrift RPC Server for JDBC/ODBC connectivity! • Hive on Spark! – Cloudera project! • Spork! – Pig on Spark! – Collabortion: Sigmoid Analytics and Twitter! • Parquet! – Columnar File! – Collaboration: Twitter and Cloudera (Impala)!
- 20. Spark Use Cases Spark SQL! • Ad hoc, exploratory, interactive analytics! • Traditional ETL, analytics, and reporting! Spark Streaming + MLlib! • Real-time traffic shifting based on CDN performance (video or music streaming)! • Proactively notify customers when errors are elevated (SaaS model)! MLlib + GraphX! • Recommendations, anomaly/spam detection, related news! • Social analytics: “You May Know”, top influencers! Spark Streaming + Spark SQL! • Lambda Architecture: batch + real-time!
- 21. Explosion of Specialized Systems
- 23. RDD Overview • Core Spark abstraction! • Represents partitions ! across the cluster nodes! • Enables parallel processing ! on data sets! • Partitions can be in-memory! or on-disk! • Immutable, recomputable, ! fault tolerant! • Contains transformation history (“lineage”) for whole data set!
- 24. RDD Interface Lineage (Required)! 1. Set of partitions! 2. List of dependencies (0, 1, many)! 3. Compute function per partition! Execution optimizations (Optional)! 4. Partitioner (ie. Hash, Range)! 5. Preferred locations per partition!
- 25. Types of RDDs • Spark Out-of-the-Box! – HadoopRDD – FilteredRDD! – MappedRDD! – PairRDD! – ShuffledRDD! – UnionRDD! – DoubleRDD! – JdbcRDD! – JsonRDD! – SchemaRDD! – VertexRDD! – EdgeRDD! • External! – CassandraRDD (DataStax)! – GeoRDD (Esri)! – EsSpark (ElasticSearch)!
- 26. Hadoop RDD Lineage (Required)! 1. Set of partitions! 1 per HDFS block! 2. List of dependencies! None! 3. Compute function ! per partition! Read HDFS block ! Execution optimizations ! (Optional)! 4. Partitioner! (ie. Hash, Range)! None! 5. Preferred locations ! per partition! HDFS block location ! from NameNode!
- 28. Filtered RDD Lineage (Required)! 1. Set of partitions! Same as parent RDD! 2. List of dependencies! Parent RDD! 3. Compute function! Compute parent and filter! Execution optimizations ! (Optional)! 4. Partitioner ! (ie. Hash, Range)! None! 5. Preferred locations ! per partition! Same as parent RDD!
- 29. Mapped RDD Lineage (Required)! 1. Set of partitions! Same as parent RDD! 2. List of dependencies! Parent RDD! 3. Compute function! Compute parent, ! apply map function! Execution optimizations ! (Optional)! 4. Partitioner ! (ie. Hash, Range)! None! 5. Preferred locations ! per partition! Same as parent RDD!
- 30. Joined RDD Lineage (Required)! 1. Set of partitions! One per reduce task ! (PART- file)! 2. List of dependencies! Every parent ! (many to many shuffle)! 3. Compute function! Read and join ! shuffled data! Execution optimizations ! (Optional)! 4. Partitioner ! (ie. Hash, Range)! HashPartitioner! 5. Preferred locations! per partition! None!
- 31. RDD Operator DAG
- 32. RDD Operator DAG
- 33. Demo: Joining Users and Gender • Load user data! • Load gender data! • Join user data ! with gender data! • Analyze lineage
- 34. Join Optimizations • When joining large dataset with small dataset (reference data)! • Broadcast small dataset to each node/partition of large dataset (one broadcast per worker node)!
- 35. Spark Core API
- 36. Spark Core API Overview • Richer, more expressive than MapReduce! • “Distributed Scala collections”! • Native support for Java, Scala, Python, SQL, and R* (coming soon)! • Unified API across all libraries! • Operations ! – Transformations (lazy evaluation DAG)! – Actions (execute transformations)!
- 37. Why Scala? • Java/JVM is used by all popular big data frameworks! – Lucene, SOLR, Hadoop, Mahout, Flume, Cassandra, HBase, Giraph, Storm, Avro, Protocol Buffers, Thrift, Parquet, etc! • Supports closures! – Allows shipping code to data! • Functional, but not obscure! • Typing allows performance optimizations! • Leverage Spark REPL! • It’s cool!!
- 41. Job Scheduler • Job! – Contains many stages! • Stages! – Contains many tasks! • Job resource allocation is aggressive! – FIFO! • Long-running jobs may starve resources for other jobs! – Fair! • Round-robin to prevent resource starvation! • Fair Scheduler Pools! – High-priority pool for important jobs! – Separate user pools! – FIFO within the pool! – Modeled after Hadoop Fair Scheduler!
- 42. Spark Execution Model Overview • Parallel, distributed! • DAG-based, lazy evaluation! • Allows optimizations! – Reduce disk I/O! – Reduce shuffle I/O! – Single pass through dataset! – Parallel execution! – Task pipelining! • Data locality and rack awareness! • PROCESS, NODE, RACK, ANY! • Fault tolerance: RDD lineage graphs!
- 44. Task Pipelining
- 45. Daytona GraySort Contest 250,000 partitions, on-disk (No caching)!
- 46. Improved Shuffle & Sort • aka “Sort-based Shuffle”! • Released across both Spark 1.1 and 1.2! • Improves performance for large number of partitions/reducers (>10,000)! • Reduces memory pressure, GC, and OOMs! • Reduces OS open files! • Sequential versus random disk I/O! • Saturate the network link! – ~125 MB/s (1 GB ethernet)! – 1.25 GB/s (10 GB ethernet)!
- 47. Improved Shuffle • Sort-based Shuffle (SPARK-2045)! – 1 file for all partitions (grouped by partition key)! • Versus many files (1 per partition key)! – Sorted in-memory, sequential file write I/O! • Versus random file write I/O (1 per partition key)! • Java-Netty Network Module (SPARK-2468)! – Replaces old nio socket-level code! – epoll, zero-copy, kernel-space! – Self-managed memory pool, less GC! • External Shuffle Service (SPARK-3796)! – Decoupled from the Executor JVMs! – Uses Network Module! – Can serve shuffle files after an Executor crash or during GC!
- 48. Improved Sort • TimSort! – Combo of merge and insertion sort! – Optimized for partially-ordered datasets! – Used in both map and reduce phases! • Reduced Data Movement, Cache Misses! – Sort using (key, pointer) pairs! – GraySort keys are 10 bytes (2^3.16227766??)! – Compressed Oops = 4 bytes! – 10 key bytes + 2 pad bytes + 4 pointer bytes = 16 bytes (2^4!!)! – AlphaSort Paper ~1995 (Chris Nyberg, Jim Gray)!
- 49. Elastic Spark • Scaling up/down based on resource idle threshold! • Common problem! – Long-running apps hold resources even when idle! – Spark Shell is a long-running Spark app! – Spark Streaming is a long running Spark app! • Best practice! – Scale up quickly (exponentionall)! – Scale down slowly!
- 51. Cluster Resource Managers • Spark Standalone (AMPLab, Spark default)! – Suitable for a lot of production workloads! – Only Spark workloads! • YARN (Hadoop)! – Allows hierarchies of resources! – Kerberos integration! – Multiple workloads from different execution frameworks! • Hive, Pig, Spark, MapReduce, Cascading, etc! • Mesos (AMPLab)! – Similar to YARN, but allows elastic allocation to disparate execution frameworks! – Coarse-grained ! • Single, long-running Mesos tasks runs Spark mini tasks! – Fine-grained! • New Mesos task for each Spark task! • Higher overhead, not good for long-running Spark jobs (ie. Streaming)!
- 54. Spark Master High Availability • Multiple Master Nodes! • ZooKeeper maintains current Master! • Existing applications and workers will be notified of new Master election • New applications and workers need to explicitly specify current Master! • Alternatives (Not recommended)! – Local filesystem ! – NFS Mount!
- 56. Spark SQL
- 57. HiveQL (v13)
- 58. Hive In-Memory Caching • In-memory columnar storage format! • Physically stores columns together (instead of rows)! • Optimized for fast column-based aggregations! – SELECT AVG(age) FROM users! – Not “SELECT *” which is row-based! • Predicate pushdowns (WHERE age >= 21)! • Column filtering (SELECT age FROM users)! • Partition pruning (WHERE date=“2015-01”)!
- 59. Parquet File Format • On-disk columnar storage format! • Heavy compression:! – Run Length Encoding! – Dictionary Encoding! – Bit packing! • Nested structures! • Self-describing, ! evolving schema! • Optimized for:! – Column-based ! aggregations! – Data at rest! • Not optimized for:! – SELECT * (row-based)! – INSERTS/UPDATES (entire dataset needs re-encoding)!
- 60. Demo: Rating Aggregations (SparkSQL, Parquet) • Load CSV! • Convert to JSON! • Query JSON using SQL (slow)! • Convert to Parquet file! • Save Parquet file to disk! • Query Parquet using SQL (fast!)
- 61. MLlib and GraphX! (Machine Learning and! Graph Processing)
- 62. Machine Learning Algos • Supervised! – Prediction: Train a model with existing data + label, predict label for new data! • Classification (categorical or numeric)! • Regression (continuous numeric)! – Recommendation: recommend to similars! • User-user, item-item, user-item similarity! • Unsupervised! – Clustering: Find natural clusters in data based on similarities (No training needed)! • Must be parallelizable!!
- 63. ML Model Training • Input Data (Instances)! – Training data (~80%)! – Hold-out data: model validation (~10%) & testing (~10%)! – Shape/density: Tall & skinny, short & wide, dense vs. sparse?! • Features! – Columns/attributes of instances! – Types: Categorical, numeric (categorical & continuous)! – Recommender: user id, product id! – Clustering: height, religion, interest! – Combine features to derive new ones! – Requires domain knowledge! • Labels! – Supervised: predictions, classifications! • Model Hyperparameters! – Convergence threshold, number of iterations, etc!
- 64. ML Model Validation & Testing • Use hold-out data (20%)! – Compare known validation label to model-predicted label! • Measure the difference with evaluation criteria:! – Root Mean Squared Error (RMSE)! – Mean Absolute Error (MAE)! • Optimize evaluation criteria! – Choose the features and hyperparameters with the best evaluation criteria! – ML Optimizer/ML Pipelines (AMPLab)! • K-folds validation! – Split input data k-ways! – Alternate the role of each split: training, validation, testing! – Especially useful for smaller datasets! • Avoid overfitting to a particular dataset!!
- 65. Shape and Density of Input • Convert data to matrix! – Rows (instances) x Columns (features)! • SparseVector vs. DenseVector! – Sparse: O(12nnz+4); ! nnz = number of non-zeros ! – Dense requires O(8n); ! n = size of vector! • SparseVector: wide rows! – Lots of 0’s (most large distributed datasets)! – LIBSVM Format! • Some ML algos optimized for tall and skinny, some for short and wide! – Depends on parallelism/shuffle characteristics!
- 66. Categorical Features! • ML algos require data to be numeric (Double)! • Categorical is usually not numeric! – ie. NFL Teams: Bears, 49’ers, Steelers! • Index? 1->Bears 2->49’ers 3-> Steelers! – No! ML algo will assign irrelevant meaning to number! • Instead, use “one-hot” or “1 of n” encoding! – Convert categories to unique combos of 0’s and 1’s! – ie. Bears -> 1,0,0! 49’ers -> 0,1,0! Steelers-> 0,0,1! • Lots of open Spark PR’s for one-hot impls!
- 67. Similarity Algos (1/2) • DIMSUM! – “Dimension Independent Matrix Square Using MapReduce”! – From Twitter! – Merged into MLlib as RowMatrix.columnSimilarities()! – Similar to Cosine Similarity, but overcomes O(m*n^2) shuffle by sampling! – Becomes O(n^2/Epsilon^2); n = columns, Epsilon is threshold! – Sampling requires ! shuffling! – 40% improvement in ! Twitter Ad similarity ! computation h#ps://blog.twi#er.com/2014/all-‐pairs-‐similarity-‐via-‐dimsum
- 68. Similarity Algos (2/2) • LSH ! – Locality-Sensitive Hash! – Hashes values and creates buckets of similar values! – Hash function takes into account similarity using euclidean distance! – Requires pre-processing to create buckets! – Could suffer from skew! • Reduce the space and return approximation! – O(n^2) brute force w/o LSH! – O(n/m * m^2) w/ LSH!
- 69. Common LSH Use Case • ”I have a 500k x 500k matrix and I want to calculate pair-wise similarities”! – O(n^2), cartesian, lots of shuffle!! • Use LSH!! – Divide similar data into ! m buckets! – Compare between buckets! – Reduces down to ! O(n/m * m^2); ! where m = # of buckets! • O(250,000,000,000) vs.! O(25,000,000); m=50!
- 70. Serving Model Results • Cold Start Problem! – New users suffer cold start! – Need suitable starting points! • Top-rated users from Spark SQL! • Most-desirable users from PageRank! • Randomization! – Don’t underestimate the power of randomizing! – Need to surface results that wouldn’t otherwise surface! • A/B test diff models to determine actual effectiveness! • Combine model results! – Track which model the result came from! – Remember that position in results matters!
- 71. Demo: Generating Matches ! (PageRank & ALS) GraphX: PageRank (TopK) MLlib: ALS (Recommend)
- 72. Spark Streaming
- 73. “Streaming” Kinesis Streaming Video ! Streaming Piping Big Data! Streaming
- 74. Spark Streaming Overview! • Low latency, high throughput, fault- tolerance (mostly)! • Long-running Spark application! • Supports Flume, Kafka, Twitter, Kinesis, Socket, File, etc.! • Graceful shutdown, in-flight message draining! • Uses Spark Core, DAG Execution Model, and Fault Tolerance! • Submits micro-batch jobs to the cluster!
- 76. Spark Streaming Use Cases • ETL and enrichment ! of streaming data! on injest! – Spark Streaming,! Spark SQL, MLlib! • Lambda Architecture! – Spark Streaming, ! Spark SQL, MLlib, GraphX! – Everything!!! ! • Operational dashboards! – Lambda Architecture! – Spark Streaming, ! Spark SQL!
- 77. Discretized Stream (DStream) • Core Spark Streaming abstraction! • Micro-batches of RDDs! • Operations similar to RDD! • Operates on all underlying RDDs of DStream!
- 78. Spark Streaming API Overview • Rich, expressive API similar to core! • Operations! – Transformations (lazy)! – Actions (execute transformations)! • Window and Stateful Operations! – Current state depends on previous state! • Requires 2 types of durable checkpointing (HDFS) to snip long-running lineages! – Metadata: DStream DAG, queue of jobs, ! – Data: input data, window and stateful data! • Register DStream as a Spark SQL table for querying?! Wow.
- 80. DStream Actions
- 81. Window and State DStream Operations
- 82. DStream Window and State Example
- 83. Spark Streaming ! Cluster Deployment
- 85. Adding Receivers
- 86. Adding Workers
- 88. Spark Streaming + Kinesis Architecture
- 90. Demo: Real-time ALS Matches Kinesis Streaming Scala: …/scala/org/apache/spark/examples/streaming/KinesisWordCountASL.scala Java: …/java/org/apache/spark/examples/streaming/JavaKinesisWordCountASL.java h#ps://github.com/apache/spark/blob/master/extras/kinesis-‐asl/src/main/…
- 91. Kinesis Best Practices • Avoid resharding! – Over-provision shards to avoid splitting/merging & increase throughput! – Remember to include egress rate when provisioning – not just ingest! – Choose high-cardinality partition hash key to avoid hot spots/shards (ie. CustomerID, etc)! • Avoid excess PUTs! – Batch vs. single records! – Kinesis Mobile Client supports batch to reduce battery consumption! – Increases throughput! – Filter on client-side! – Async producer like FluentD, Log4J, Flume! • Avoid data loss! – Use an archiver consumer app to move the data into S3 for longer-term durability past 24 hours! – Write to durable storage (Spark 1.2 WAL) before checkpointing! – De-dupe as deep into the pipeline as possible – allows retry! – Use deterministic checkpoint strategy to get repeatable failover results ! (ie. number of records since last checkpoint)! – Idempotent processing where possible!
- 93. Types of Streaming Sources • Batched! – Flume, Kafka, Kinesis! – Improves throughput at expense of possible duplication during failure replay! • Partitioned! – Kafka, Kinesis! – Different receivers dedicated to different partitions! – Improves throughput through parallelism! – Order is only guaranteed within a single partition! • Buffered! – Flume, Kafka, Kinesis! – Allows back pressure, rate limiting, etc! • Reliable (Checkpointed)! – Kafka, Kinesis! – Allows replay from specific checkpoint! – Requires ACK after each batch is durably stored within Spark Streaming! • Transactional! – Flume! – Start, Commit semantics!
- 94. Message Delivery Guarantees • Exactly once [1]! – No loss! – No redeliver! – Perfect delivery! – Incurs higher latency for transactional semantics! – Spark default per batch using DStream lineage! – Degrades to less guarantees depending on source! • At least once [1..n]! – No loss! – Possible redeliver! • At most once [0,1]! – Possible loss! – No redeliver! – *Best configuration if some data loss is acceptable! • Ordered! – Per partition: Kafka, Kinesis! – Global across all partitions: Hard to scale!
- 95. Types of Checkpoints Spark! 1. Metadata: Spark checkpointing of StreamingContext DStreams and metadata ! 2. Data: Lineage of input data, state, and window DStream operations! Kinesis! 3. Kinesis Client Library (KCL) checkpoints current position within shard! – Checkpoint info is stored in DynamoDB per Kinesis application keyed by shard!
- 96. Fault Tolerance • Points of Failure! – Receiver! • Should restart, replace automatically! • Possible data loss if not using Write Ahead Log (WAL)! – Driver! • Application is torn down including worker allocations! – Worker/Executor! • Normal Spark partition-level recovery occurs! • Possible Solutions! – Use HDFS File Source for durability – Data Replication! – Secondary/Backup Nodes! – Checkpoints! • Stream, Window, and State info!
- 97. Streaming Receiver Failure • Use multiple receivers pulling from multiple shards! • Use checkpoint-enabled, reliable, sharded source (ie. Kafka, Kinesis) for replay! • Data is replicated to 1 other node in-memory immediately upon ingestion! – 2nd node is chosen at random! – Spills to disk if data doesn’t fit in memory! • Write Ahead Log (WAL)! – Prevents losing in-flight metadata and data (received, but not processed)! – Stored in HDFS! – ACK back to source when safely stored in WAL! – Spark processes the stored-but-not-processed data upon failover!
- 98. Streaming Driver Failure • Use a backup Driver ! – Use DStream metadata checkpoint info to recover! – StreamingContext.getOrCreate()! • Single point of failure! – Interrupts stream processing! – State and Window data is lost! – Use checkpointing to avoid data loss! • Streaming Driver is a long-running Spark application! – Schedules long-running receivers! – Checkpointing prevents infinite DStream lineage for long-running streaming app!
- 99. Stream Worker/Processor Failure • No problem!! • DStream RDD partitions will be recalculated from lineage! • Causes temporary processing blip during failover!
- 100. Spark Streaming! Monitoring and Tuning
- 101. Monitoring • Monitor driver, receiver, worker nodes, and streams! • Alert upon failure or unusually high latency! • Spark Web UI! – Streaming tab! • Ganglia, CloudWatch (AWS)! • Custom StreamingListener callback hooks during DStream processing!
- 102. Spark Web UI
- 103. Tuning (1/2) • DStream batch time interval! – Higher: reduces overhead by submitting less tasks overall! – Lower: keeps latencies low! – Sweet spot: DStream job time (scheduling + processing) is steady and less than batch interval! ! • Checkpoint time interval! – Higher: reduces load on checkpoint overhead! – Lower: reduces amount of data loss on failure! – Recommendation: 5-10x sliding window interval!
- 104. Tuning (2/2) • Use DStream.repartition() ! – Increase parallelism of processing DStream jobs across more cluster nodes! • Use spark.streaming.unpersist=true! – Let’s the Streaming Framework figure out when to unpersist out-of-scope resources! • Use Concurrent Mark and Sweep (CMS) GC Strategy! – For consistent processing times! – Note: Default stores data as serialized byte arrays to minimize GC! • Consistency is key!!!
- 105. Spark and the Lambda Architecture
- 106. Lambda Architecture Overview • Batch Layer! – Immutable, ! Batch read,! Append-only write! – Source of truth! – ie. HDFS! • Speed Layer! – Mutable,! Random read/write! – Most complex! – Recent data only! – ie. Cassandra! • Serving Layer! – Immutable,! Random read, ! Batch write! – ie. ElephantDB!
- 107. Spark + AWS + Lambda
- 108. Spark + Lambda + GraphX + MLlib
- 109. Approximations
- 110. Approximation Overview • Required for scaling efficiently! • Speed up analysis of large datasets! • Reduce size of working dataset! • Data is messy! • Collection of data is messy! • Exact isn’t always necessary! • “Approximate is the new Exact”! !
- 111. Probabilistic Data Structs/Algos • HyperLogLog! – Count distinct elements for a date range! – 16K for 99% accuracy! • Count-min Sketch! – Heavy hitters, top-k, most-frequent elements! – Supports streaming data! – Fixed amount of space for all counts! • Bloom Filter! – Search/filter elements! – Bitset representation of hash! – “Maybe yes, definitely no”!
- 112. Benefits of Approximations Figure: Memory Savings with ApproximaLon Techniques (h#p://highlyscalable.wordpress.com/2012/05/01/probabilisLc-‐structures-‐web-‐analyLcs-‐data-‐mining/)
- 113. Spark’s Approximation Methods • BlinkDB! – Approximate error-bound, time-bound “online aggregation” queries! • RDD: sample(), takeSample()! – Bernouilli and Poisson Sampling! • PairRDD: countApproxDistinctByKey()! – HyperLogLog • Spark Streaming and Twitter Algebird! • Count-min Sketch! • MLlib RowMatrix.columnSimilarities()! – DIMSUM approximation! • Bloom Filters! – Everywhere!!
- 114. Spark’s Statistics Library • Correlations! – Dependence between 2 random variables! – Pearson, Spearman! • Hypothesis Testing! – Measure of statistical significance! – Chi-squared test! • Stratified Sampling! – Sample separately from different sub-populations! – Bernoulli and Poisson sampling! – With and without replacement! • Random data generator! – Uniform, standard normal, and Poisson distribution
- 116. Summary • Spark, Spark Streaming Overview ! • Use Cases! • API and Libraries! • Machine Learning! • Graph Processing! • Execution Model! • Fault Tolerance! • Cluster Deployment! • Monitoring! • Scaling and Tuning! • Lambda Architecture! • Probabilistic Data Structs/Algos! • Approximations! http://sparkinaction.com! http://effectivespark.com Thanks!!! ! Chris Fregly! @cfregly!