SF Big Analytics: Introduction to Succinct by UC Berkeley AmpLab
2 likes384 views
The document discusses techniques for implementing interactive queries at scale, specifically addressing performance degradation issues when data size exceeds memory limits. It introduces a solution called Succinct, which allows for compressed data representation and facilitates complex queries such as search, range queries, and regular expressions directly on this compressed data without the need for additional indexing. The evaluation showcased that Succinct is significantly faster and more space-efficient compared to existing systems like Apache Spark and Elasticsearch, especially in use cases involving unstructured data and regular expressions.
![Our Solution
BlowFish [NSDIā16]
Succinct [NSDIā15]
SuccinctāØ
Encryption
GraphStore
KVStore
ColumnarStore
RowStore
UnstructuredData](https://image.slidesharecdn.com/succinctmeetup-161008010628/85/SF-Big-Analytics-Introduction-to-Succinct-by-UC-Berkeley-AmpLab-19-320.jpg)
![Our Solution
ā£ Compressed representation ā More queries in faster storage
BlowFish [NSDIā16]
Succinct [NSDIā15]
SuccinctāØ
Encryption
GraphStore
KVStore
ColumnarStore
RowStore
UnstructuredData](https://image.slidesharecdn.com/succinctmeetup-161008010628/85/SF-Big-Analytics-Introduction-to-Succinct-by-UC-Berkeley-AmpLab-20-320.jpg)
![Our Solution
ā£ Compressed representation ā More queries in faster storage
ā£ Rich functionality directly on compressed representation
ā£ Search, RegEx, Range queries
BlowFish [NSDIā16]
Succinct [NSDIā15]
SuccinctāØ
Encryption
GraphStore
KVStore
ColumnarStore
RowStore
UnstructuredData](https://image.slidesharecdn.com/succinctmeetup-161008010628/85/SF-Big-Analytics-Introduction-to-Succinct-by-UC-Berkeley-AmpLab-21-320.jpg)
![Our Solution
ā£ Compressed representation ā More queries in faster storage
ā£ Rich functionality directly on compressed representation
ā£ Search, RegEx, Range queries
ā£ Flexible support for diļ¬erent data models
BlowFish [NSDIā16]
Succinct [NSDIā15]
SuccinctāØ
Encryption
GraphStore
KVStore
ColumnarStore
RowStore
UnstructuredData](https://image.slidesharecdn.com/succinctmeetup-161008010628/85/SF-Big-Analytics-Introduction-to-Succinct-by-UC-Berkeley-AmpLab-22-320.jpg)
![Our Solution
ā£ Compressed representation ā More queries in faster storage
ā£ Rich functionality directly on compressed representation
ā£ Search, RegEx, Range queries
ā£ Flexible support for diļ¬erent data models
ā£ Handles skewed & time-varying workloads
BlowFish [NSDIā16]
Succinct [NSDIā15]
SuccinctāØ
Encryption
GraphStore
KVStore
ColumnarStore
RowStore
UnstructuredData](https://image.slidesharecdn.com/succinctmeetup-161008010628/85/SF-Big-Analytics-Introduction-to-Succinct-by-UC-Berkeley-AmpLab-23-320.jpg)
![Changes in Spatial Skew
Study on Facebook
Warehouse Cluster
[HotStorageā13]](https://image.slidesharecdn.com/succinctmeetup-161008010628/85/SF-Big-Analytics-Introduction-to-Succinct-by-UC-Berkeley-AmpLab-253-320.jpg)
![Changes in Spatial Skew
Transient failures ā 90% of failuresStudy on Facebook
Warehouse Cluster
[HotStorageā13]](https://image.slidesharecdn.com/succinctmeetup-161008010628/85/SF-Big-Analytics-Introduction-to-Succinct-by-UC-Berkeley-AmpLab-254-320.jpg)
![Changes in Spatial Skew
Transient failures ā 90% of failures
Replica creation delayed by 15 mins
Study on Facebook
Warehouse Cluster
[HotStorageā13]](https://image.slidesharecdn.com/succinctmeetup-161008010628/85/SF-Big-Analytics-Introduction-to-Succinct-by-UC-Berkeley-AmpLab-255-320.jpg)
![Changes in Spatial Skew
Transient failures ā 90% of failures
Replica creation delayed by 15 mins
Study on Facebook
Warehouse Cluster
[HotStorageā13]
Leads to variation in load over time](https://image.slidesharecdn.com/succinctmeetup-161008010628/85/SF-Big-Analytics-Introduction-to-Succinct-by-UC-Berkeley-AmpLab-256-320.jpg)
![Changes in Spatial Skew
Transient failures ā 90% of failures
Replica creation delayed by 15 mins
Replica#1
Replica#2
Replica#3
Data Partitions Request Queues
Study on Facebook
Warehouse Cluster
[HotStorageā13]
Leads to variation in load over time](https://image.slidesharecdn.com/succinctmeetup-161008010628/85/SF-Big-Analytics-Introduction-to-Succinct-by-UC-Berkeley-AmpLab-257-320.jpg)
![Changes in Spatial Skew
Transient failures ā 90% of failures
Replica creation delayed by 15 mins
Replica#1
Replica#2
Replica#3
Data Partitions Request Queues
Study on Facebook
Warehouse Cluster
[HotStorageā13]
Leads to variation in load over time](https://image.slidesharecdn.com/succinctmeetup-161008010628/85/SF-Big-Analytics-Introduction-to-Succinct-by-UC-Berkeley-AmpLab-258-320.jpg)
![Changes in Spatial Skew
Transient failures ā 90% of failures
Replica creation delayed by 15 mins
Replica#1
Replica#2
Replica#3
Data Partitions Request Queues
Study on Facebook
Warehouse Cluster
[HotStorageā13]
Leads to variation in load over time](https://image.slidesharecdn.com/succinctmeetup-161008010628/85/SF-Big-Analytics-Introduction-to-Succinct-by-UC-Berkeley-AmpLab-259-320.jpg)
SF Big Analytics: Introduction to Succinct by UC Berkeley AmpLab
- 1. Succinct: Fast Interactive Queries Anurag Khandelwal
- 2. Interactive Queries at Scale
- 3. Interactive Queries at Scale Search Tweets by @AMPLab about #Succinct
- 4. Interactive Queries at Scale Search Regular Expressions Tweets by @AMPLab about #Succinct Links to Berkeley or Stanford domainsāØ .*(berkeley|stanford).edu
- 5. Interactive Queries at Scale Search Regular Expressions Range Queries Tweets by @AMPLab about #Succinct Links to Berkeley or Stanford domainsāØ .*(berkeley|stanford).edu All my Facebook posts between 2013 and 2016
- 6. Interactive Queries at Scale Search Regular Expressions Range Queries Graph Queries Tweets by @AMPLab about #Succinct Links to Berkeley or Stanford domainsāØ .*(berkeley|stanford).edu All my Facebook posts between 2013 and 2016 Friends of my friends who like trekking
- 7. Interactive Queries at Scale Search Random Access Regular Expressions Range Queries Graph Queries Aggregate Queries Updates Tweets by @AMPLab about #Succinct Links to Berkeley or Stanford domainsāØ .*(berkeley|stanford).edu All my Facebook posts between 2013 and 2016 Friends of my friends who like trekking
- 8. Interactive Queries at Scale Search Random Access Regular Expressions Range Queries Graph Queries Aggregate Queries Updates Compute Platforms
- 9. Interactive Queries at Scale Search Random Access Regular Expressions Range Queries Graph Queries Aggregate Queries Updates Compute Platforms Query Engines
- 10. Interactive Queries at Scale Search Random Access Regular Expressions Range Queries Graph Queries Aggregate Queries Updates Compute Platforms Query Engines Data Stores
- 11. Interactive Queries at Scale
- 12. Interactive Queries at Scale Todayās focus on two main issues:
- 13. Interactive Queries at Scale ā£ Performance degradation when data size > memory Todayās focus on two main issues:
- 14. Interactive Queries at Scale ā£ Performance degradation when data size > memory Todayās focus on two main issues: Throughput (Ops) 0 500 1000 1500 2000 Input Size 1GB 2GB 4GB 8GB 16GB 32GB 64GB 128GB
- 15. Interactive Queries at Scale ā£ Performance degradation when data size > memory Todayās focus on two main issues: Throughput (Ops) 0 500 1000 1500 2000 Input Size 1GB 2GB 4GB 8GB 16GB 32GB 64GB 128GB
- 16. Interactive Queries at Scale ā£ Performance degradation when data size > memory Todayās focus on two main issues: ā£ Handling skewed query workloads Throughput (Ops) 0 500 1000 1500 2000 Input Size 1GB 2GB 4GB 8GB 16GB 32GB 64GB 128GB
- 17. Interactive Queries at Scale ā£ Performance degradation when data size > memory Todayās focus on two main issues: ā£ Handling skewed query workloads Throughput (Ops) 0 500 1000 1500 2000 Input Size 1GB 2GB 4GB 8GB 16GB 32GB 64GB 128GB
- 18. Interactive Queries at Scale ā£ Performance degradation when data size > memory Todayās focus on two main issues: ā£ Handling skewed query workloads Throughput (Ops) 0 500 1000 1500 2000 Input Size 1GB 2GB 4GB 8GB 16GB 32GB 64GB 128GB Maximum sustainable throughput
- 19. Our Solution BlowFish [NSDIā16] Succinct [NSDIā15] SuccinctāØ Encryption GraphStore KVStore ColumnarStore RowStore UnstructuredData
- 20. Our Solution ā£ Compressed representation ā More queries in faster storage BlowFish [NSDIā16] Succinct [NSDIā15] SuccinctāØ Encryption GraphStore KVStore ColumnarStore RowStore UnstructuredData
- 21. Our Solution ā£ Compressed representation ā More queries in faster storage ā£ Rich functionality directly on compressed representation ā£ Search, RegEx, Range queries BlowFish [NSDIā16] Succinct [NSDIā15] SuccinctāØ Encryption GraphStore KVStore ColumnarStore RowStore UnstructuredData
- 22. Our Solution ā£ Compressed representation ā More queries in faster storage ā£ Rich functionality directly on compressed representation ā£ Search, RegEx, Range queries ā£ Flexible support for diļ¬erent data models BlowFish [NSDIā16] Succinct [NSDIā15] SuccinctāØ Encryption GraphStore KVStore ColumnarStore RowStore UnstructuredData
- 23. Our Solution ā£ Compressed representation ā More queries in faster storage ā£ Rich functionality directly on compressed representation ā£ Search, RegEx, Range queries ā£ Flexible support for diļ¬erent data models ā£ Handles skewed & time-varying workloads BlowFish [NSDIā16] Succinct [NSDIā15] SuccinctāØ Encryption GraphStore KVStore ColumnarStore RowStore UnstructuredData
- 25. Existing Techniques Data Scans SEARCH( ) Example:
- 26. Existing Techniques Data Scans SEARCH( ) Example: Ex: Apache Spark
- 27. Existing Techniques Data Scans SEARCH( ) Example: Ex: Apache Spark
- 28. Existing Techniques Data Scans SEARCH( ) Example: Ex: Apache Spark
- 29. Existing Techniques Data Scans Low storage High Latency SEARCH( ) Example: Ex: Apache Spark
- 30. Existing Techniques Data Scans Indexes Low storage High Latency SEARCH( ) Example: Ex: Apache Spark
- 31. Existing Techniques Data Scans Indexes Low storage High Latency SEARCH( ) Example: Ex: Apache Spark Ex: SOLR
- 32. Existing Techniques Data Scans Indexes Low storage High Latency SEARCH( ) Example: Ex: Apache Spark Ex: SOLR
- 33. Existing Techniques 0, 10, 14, 16, 19, 26, 29 1, 4, 5, 8, 20, 22, 24 2, 15, 17, 27 3, 6, 7, 9, 12, 13, 18, 23 .. 11, 21 Data Scans Indexes Low storage High Latency SEARCH( ) Example: Ex: Apache Spark Ex: SOLR
- 34. Existing Techniques 0, 10, 14, 16, 19, 26, 29 1, 4, 5, 8, 20, 22, 24 2, 15, 17, 27 3, 6, 7, 9, 12, 13, 18, 23 .. 11, 21 Data Scans Indexes Low storage High Latency SEARCH( ) Example: Ex: Apache Spark Ex: SOLR
- 35. Existing Techniques 0, 10, 14, 16, 19, 26, 29 1, 4, 5, 8, 20, 22, 24 2, 15, 17, 27 3, 6, 7, 9, 12, 13, 18, 23 .. 11, 21 Data Scans Indexes Low storage High Latency High storage Low Latency SEARCH( ) Example: Ex: Apache Spark Ex: SOLR
- 36. Succinct
- 39. Succinct Queries executed directly on the compressed representation Succinct
- 40. Succinct Queries executed directly on the compressed representation Low Storage Low Latency Succinct
- 41. Succinct Queries executed directly on the compressed representation Low Storage Low Latency Succinct What makes Succinct unique
- 42. Succinct Queries executed directly on the compressed representation Low Storage Low Latency Succinct What makes Succinct unique No additional indexes Query responses embedded in the compressed representation
- 43. Succinct Queries executed directly on the compressed representation Low Storage Low Latency Succinct What makes Succinct unique No additional indexes Query responses embedded in the compressed representation No data scans Functionality of indexes
- 44. Succinct Queries executed directly on the compressed representation Low Storage Low Latency Succinct What makes Succinct unique No additional indexes Query responses embedded in the compressed representation No data scans Functionality of indexes No decompression Queries directly on the compressed representation (except for data access queries)
- 45. Succinct Queries executed directly on the compressed representation Low Storage Low Latency Succinct
- 46. Succinct Queries executed directly on the compressed representation Low Storage Low Latency Succinct Scale In-memory data sizes >= memory capacity
- 47. Succinct Queries executed directly on the compressed representation Low Storage Low Latency Succinct Scale In-memory data sizes >= memory capacity Complex queries Search, range, random access, RegEx
- 48. Succinct Queries executed directly on the compressed representation Low Storage Low Latency Succinct Scale In-memory data sizes >= memory capacity Complex queries Search, range, random access, RegEx Interactivity Avoids data scans and decompression
- 50. Succinct Data Representation Builds on a large body of theory work
- 51. Succinct Data Representation Builds on a large body of theory work Suļ¬x Arrays
- 52. Succinct Data Representation Builds on a large body of theory work Suļ¬x Arrays ā£ Strong functionality (search)
- 53. Succinct Data Representation Builds on a large body of theory work Suļ¬x Arrays ā£ Strong functionality (search) ā£ No structure
- 54. Succinct Data Representation Builds on a large body of theory work Suļ¬x Arrays ā£ Strong functionality (search) ā£ No structure Compression?
- 55. Succinct Data Representation Builds on a large body of theory work Suļ¬x Arrays ā£ Strong functionality (search) ā£ No structure Compression? ā£ Sample the suļ¬x array
- 56. Succinct Data Representation Builds on a large body of theory work Suļ¬x Arrays ā£ Strong functionality (search) ā£ No structure Compression? ā£ Sample the suļ¬x array ā£ Store set of pointers to compute unsampled values on the ļ¬y
- 57. Succinct Data Representation Builds on a large body of theory work Suļ¬x Arrays ā£ Strong functionality (search) ā£ No structure Compression? ā£ Sample the suļ¬x array ā£ Store set of pointers to compute unsampled values on the ļ¬y Possesses structure that enables compression!
- 59. Succinct Data Model ā£ Unstructured data ā£ Key-value stores (Voldemort, Dynamo) ā£ Document store (Elasticsearch, MongoDB) ā£ Tables (Cassandra, BigTable) ā£ And many more .... Uniļ¬ed Interface
- 60. Succinct Data Model ā£ Unstructured data ā£ Key-value stores (Voldemort, Dynamo) ā£ Document store (Elasticsearch, MongoDB) ā£ Tables (Cassandra, BigTable) ā£ And many more .... Uniļ¬ed Interface With all the powerful queries on values, documents, columns
- 61. Data Model & Functionality For unstructured data:
- 62. Data Model & Functionality Original Input Succinct For unstructured data:
- 63. Data Model & Functionality Original Input Succinct SEARCH( )= {0, 10, 14, 16, 19, 26, 29} Search: returns oļ¬sets of arbitrary strings in uncompressed ļ¬le For unstructured data:
- 64. Data Model & Functionality Original Input Succinct SEARCH( )= {0, 10, 14, 16, 19, 26, 29} For unstructured data: Extract(0, 5) = { , , , , } Extract: returns data at arbitrary oļ¬sets in uncompressed ļ¬le
- 65. Data Model & Functionality Original Input Succinct SEARCH( )= {0, 10, 14, 16, 19, 26, 29} For unstructured data: Extract(0, 5) = { , , , , } COUNT( ) = 7 Count: returns count of arbitrary strings in uncompressed ļ¬le
- 66. Data Model & Functionality Original Input Succinct SEARCH( )= {0, 10, 14, 16, 19, 26, 29} For unstructured data: Extract(0, 5) = { , , , , } COUNT( ) = 7 Append( , , , , ) Append: appends arbitrary strings to uncompressed ļ¬le
- 67. Data Model & Functionality Original Input Succinct SEARCH( )= {0, 10, 14, 16, 19, 26, 29} For unstructured data: Extract(0, 5) = { , , , , } COUNT( ) = 7 Append( , , , , ) Range Queries, REGULAR EXPRESSIONS
- 68. Unifying the Data Models
- 69. Unifying the Data Models
- 70. Unifying the Data Models
- 71. Unifying the Data Models
- 72. Unifying the Data Models SEARCH(Column1, )
- 73. Unifying the Data Models SEARCH(Column1, )SEARCH( )
- 82. Succinct on āØ Apache Spark
- 83. Queries on Compressed RDDs
- 84. Queries on Compressed RDDs New Functionalities Document store, āØ Key-Value store search on documents, values
- 85. Queries on Compressed RDDs New Functionalities Document store, āØ Key-Value store search on documents, values Faster operations on RDDs random access, ļ¬lters avoid scans
- 86. Queries on Compressed RDDs New Functionalities Document store, āØ Key-Value store search on documents, values Faster operations on RDDs random access, ļ¬lters avoid scans More in-memory Compressed RDDs no decompression overheads
- 87. Unstructured data using SuccinctRDD
- 88. import edu.berkeley.cs.succinct._ Import classes Unstructured data using SuccinctRDD
- 89. import edu.berkeley.cs.succinct._ val rdd = ctx.textFile(ā¦).map(_.getBytes) val succinctRDD = rdd.succinct Load data & compress using Succinct Unstructured data using SuccinctRDD
- 90. import edu.berkeley.cs.succinct._ val rdd = ctx.textFile(ā¦).map(_.getBytes) val succinctRDD = rdd.succinct val offsets = succinctRDD.search("Berkeley") Find all occurrences of āBerkeleyā Unstructured data using SuccinctRDD
- 91. import edu.berkeley.cs.succinct._ val rdd = ctx.textFile(ā¦).map(_.getBytes) val succinctRDD = rdd.succinct val count = succinctRDD.count("Berkeley") val offsets = succinctRDD.search("Berkeley") Count #occurrences of āBerkeleyā Unstructured data using SuccinctRDD
- 92. import edu.berkeley.cs.succinct._ val rdd = ctx.textFile(ā¦).map(_.getBytes) val succinctRDD = rdd.succinct val bytes = succinctRDD.extract(50, 100) val count = succinctRDD.count("Berkeley") val offsets = succinctRDD.search("Berkeley") Extract 100 bytes from offset 50 Unstructured data using SuccinctRDD
- 93. Key-Value Store using SuccinctKVRDD
- 94. import edu.berkeley.cs.succinct.kv._ Import classes Key-Value Store using SuccinctKVRDD
- 95. import edu.berkeley.cs.succinct.kv._ val kvRDD = rdd.zipWithIndex.map(t => (t._2, t._1.getBytes))āØ val succinctKVRDD = kvRDD.succinctKV Load data & compress using Succinct Key-Value Store using SuccinctKVRDD
- 96. import edu.berkeley.cs.succinct.kv._ val kvRDD = rdd.zipWithIndex.map(t => (t._2, t._1.getBytes))āØ val succinctKVRDD = kvRDD.succinctKV val keys = succinctKVRDD.search("Berkeley") Find all keys for values that contain āBerkeleyā Key-Value Store using SuccinctKVRDD
- 97. import edu.berkeley.cs.succinct.kv._ val kvRDD = rdd.zipWithIndex.map(t => (t._2, t._1.getBytes))āØ val succinctKVRDD = kvRDD.succinctKV val value = succinctKVRDD.get(0) val keys = succinctKVRDD.search("Berkeley") Get value for key 0 Key-Value Store using SuccinctKVRDD
- 98. Evaluation
- 100. Evaluation Dataset Cluster Wikipedia datasetāØ ~40GB data Amazon EC2, 5 machines, 30GB RAM each
- 101. Evaluation Dataset Cluster Workload Wikipedia datasetāØ ~40GB data Amazon EC2, 5 machines, 30GB RAM each Search queries, 1-10,000 occurrences
- 102. Evaluation Dataset Cluster Workload Systems Wikipedia datasetāØ ~40GB data Amazon EC2, 5 machines, 30GB RAM each Search queries, 1-10,000 occurrences Spark, Elasticsearch
- 103. Evaluation Dataset Cluster Workload Systems Wikipedia datasetāØ ~40GB data Amazon EC2, 5 machines, 30GB RAM each Search queries, 1-10,000 occurrences Spark, Elasticsearch Caveats Absolute numbers are dataset dependent
- 104. Evaluation: Search
- 105. Evaluation: Search Takeaway: Succinct on Apache Spark is 2.5x faster than Elasticsearch while being 2.5x more space eļ¬cient.āØ (Data ļ¬ts in memory for all systems)
- 106. Support for Regular Expressions
- 107. Support for Regular Expressions Applications Data CleaningāØ Information ExtractionāØ Bioinformatics Document Stores
- 108. Support for Regular Expressions Applications Operators Data CleaningāØ Information ExtractionāØ Bioinformatics Document Stores Union, Concat, Wildcard, Repeat
- 109. Support for Regular Expressions Applications Operators Data CleaningāØ Information ExtractionāØ Bioinformatics Document Stores Union, Concat, Wildcard, Repeat Example .*(berkeley|stanford).edu
- 110. Support for Regular Expressions
- 111. Support for Regular Expressions val matches = succinctRDD.regexSearch(".*(berkeley|stanford).edu") Find all matches for the RegEx ā.*(berkeley|stanford).eduā SuccinctRDD
- 112. Support for Regular Expressions val matches = succinctRDD.regexSearch(".*(berkeley|stanford).edu") Find all matches for the RegEx ā.*(berkeley|stanford).eduā SuccinctRDD val matchKeys = succinctKVRDD.regexSearch(".*(berkeley|stanford).edu") Find all keys for values that contain the RegEx ā.*(berkeley|stanford).eduā SuccinctKVRDD
- 113. Evaluation: RegEx
- 114. Evaluation: RegEx
- 115. Evaluation: RegEx Takeaway: Succinct signiļ¬cantly speeds up RegEx queries even when all the data ļ¬ts in memory for all systems.
- 116. Succinct on Apache Spark
- 117. Succinct on Apache Spark Already in use at Elsevier Labs
- 118. Succinct on Apache Spark Already in use at Elsevier Labs ā£ Use case: Annotation Search
- 119. Succinct on Apache Spark Already in use at Elsevier Labs ā£ Use case: Annotation Search Documents
- 120. Succinct on Apache Spark Already in use at Elsevier Labs ā£ Use case: Annotation Search Documents 1, sentence, (0, 15) 2, word, (0, 4) 3, word, (5, 10) 4, word, (11, 15) Annotations
- 121. Succinct on Apache Spark Already in use at Elsevier Labs ā£ Use case: Annotation Search Documents 1, sentence, (0, 15) 2, word, (0, 4) 3, word, (5, 10) 4, word, (11, 15) Annotations āFind sentences that talk about open problems in researchā
- 122. Succinct on Apache Spark Already in use at Elsevier Labs ā£ Use case: Annotation Search Documents 1, sentence, (0, 15) 2, word, (0, 4) 3, word, (5, 10) 4, word, (11, 15) Annotations (remains|is|still) (unknown|unclear|uncertain) within <sentence> RegEx Annotation āFind sentences that talk about open problems in researchā
- 123. Succinct on Apache Spark Already in use at Elsevier Labs ā£ Use case: Annotation Search Documents 1, sentence, (0, 15) 2, word, (0, 4) 3, word, (5, 10) 4, word, (11, 15) Annotations https://spark-packages.org/package/amplab/succinct (remains|is|still) (unknown|unclear|uncertain) within <sentence> RegEx Annotation āFind sentences that talk about open problems in researchā
- 124. Problem: Skewed Query Workloads
- 125. Problem: Skewed Query Workloads Load distribution across partitions is often non-uniform
- 126. Problem: Skewed Query Workloads ā£ Succinct: Larger fraction of queries in main memory ā£ Challenge: skewed load across shards? ā£ Challenge: time varying loads? Load distribution across partitions is often non-uniform
- 127. Problem: Skewed Query Workloads ā£ Succinct: Larger fraction of queries in main memory ā£ Challenge: skewed load across shards? ā£ Challenge: time varying loads? ā£ E.g.: Memcached + MySQL deployment @ Facebook Load distribution across partitions is often non-uniform
- 128. Problem: Skewed Query Workloads Load distribution across partitions is often non-uniform
- 129. Problem: Skewed Query Workloads Load distribution across partitions is often non-uniform
- 130. Selective Replication Problem: Skewed Query Workloads Load distribution across partitions is often non-uniform Traditional approach:
- 131. Selective Replication Problem: Skewed Query Workloads Load distribution across partitions is often non-uniform #Replicas Traditional approach:
- 132. Selective Replication Problem: Skewed Query Workloads Load distribution across partitions is often non-uniform #Replicas #Replicas Ī± Load Traditional approach:
- 133. Selective Replication Problem: Skewed Query Workloads Load distribution across partitions is often non-uniform #Replicas #Replicas Ī± Load Coarse grained Traditional approach:
- 134. Selective Replication Problem: Skewed Query Workloads Load distribution across partitions is often non-uniform #Replicas #Replicas Ī± Load Coarse grained 1-2Ć throughput ā 2Ć storage Traditional approach:
- 135. Succinct + BlowFish
- 136. Succinct + BlowFish
- 137. Succinct + BlowFish
- 142. Succinct + BlowFish Storage Throughput Scans Indexes Succinct Storage-Performance tradeoļ¬ curve for each partition
- 143. Succinct + BlowFish Storage Throughput Scans Indexes Succinct Storage-Performance tradeoļ¬ curve for each partition
- 144. BlowFish: Layered Sampled Array
- 145. Recap: Succinct stores a sampled suļ¬x array BlowFish: Layered Sampled Array
- 146. Recap: Succinct stores a sampled suļ¬x array BlowFish: Layered Sampled Array Unsampled values computed on the ļ¬y
- 147. OriginalSampled āØ Array 9 15 3 0 12 8 14 5 Recap: Succinct stores a sampled suļ¬x array BlowFish: Layered Sampled Array Rate = 2 Unsampled values computed on the ļ¬y
- 148. OriginalSampled āØ Array 9 15 3 0 12 8 14 5 Recap: Succinct stores a sampled suļ¬x array BlowFish: Layered Sampled Array Rate = 2 Unsampled values computed on the ļ¬y
- 149. OriginalSampled āØ Array 9 15 3 0 12 8 14 5 9 12RATE = 8 Recap: Succinct stores a sampled suļ¬x array BlowFish: Layered Sampled Array Rate = 2 Unsampled values computed on the ļ¬y
- 150. OriginalSampled āØ Array 9 15 3 0 12 8 14 5 9 12RATE = 8 3 14RATE = 4 Recap: Succinct stores a sampled suļ¬x array BlowFish: Layered Sampled Array Rate = 2 Unsampled values computed on the ļ¬y
- 151. OriginalSampled āØ Array 9 15 3 0 12 8 14 5 9 12RATE = 8 3 14RATE = 4 15 0 8 5RATE = 2 Recap: Succinct stores a sampled suļ¬x array BlowFish: Layered Sampled Array Rate = 2 Unsampled values computed on the ļ¬y
- 152. OriginalSampled āØ Array 9 15 3 0 12 8 14 5 9 12RATE = 8 3 14RATE = 4 15 0 8 5RATE = 2 Diļ¬erent combination of layers Recap: Succinct stores a sampled suļ¬x array BlowFish: Layered Sampled Array Rate = 2 Unsampled values computed on the ļ¬y
- 153. OriginalSampled āØ Array 9 15 3 0 12 8 14 5 9 12RATE = 8 3 14RATE = 4 15 0 8 5RATE = 2 Diļ¬erent combination of layers Diļ¬erent points on tradeoļ¬ curve Recap: Succinct stores a sampled suļ¬x array BlowFish: Layered Sampled Array ā Rate = 2 Unsampled values computed on the ļ¬y
- 154. OriginalSampled āØ Array 9 15 3 0 12 8 14 5 9 12RATE = 8 3 14RATE = 4 15 0 8 5RATE = 2 Diļ¬erent combination of layers Diļ¬erent points on tradeoļ¬ curve Recap: Succinct stores a sampled suļ¬x array BlowFish: Layered Sampled Array ā Rate = 2 Layer Additions and Deletions Unsampled values computed on the ļ¬y
- 155. OriginalSampled āØ Array 9 15 3 0 12 8 14 5 9 12RATE = 8 3 14RATE = 4 15 0 8 5RATE = 2 Diļ¬erent combination of layers Diļ¬erent points on tradeoļ¬ curve Recap: Succinct stores a sampled suļ¬x array BlowFish: Layered Sampled Array ā Rate = 2 Layer Additions and Deletions Move along tradeoļ¬ curveā Unsampled values computed on the ļ¬y
- 157. BlowFish: Technical Details ā£ How should partitions share cache on a server?
- 158. BlowFish: Technical Details ā£ How should partitions share cache on a server?
- 159. BlowFish: Technical Details ā£ How should partitions share cache on a server? Low Threshold
- 160. BlowFish: Technical Details ā£ How should partitions share cache on a server? High ThresholdLow Threshold
- 161. BlowFish: Technical Details ā£ How should partitions share cache on a server? High ThresholdLow Threshold
- 162. BlowFish: Technical Details ā£ How should partitions share cache on a server? ā£ How should partitions share cache across servers? High ThresholdLow Threshold
- 163. BlowFish: Technical Details ā£ How should partitions share cache on a server? ā£ How should partitions share cache across servers? ā£ How should requests be scheduled across replicas? High ThresholdLow Threshold
- 164. BlowFish: Technical Details ā£ How should partitions share cache on a server? ā£ How should partitions share cache across servers? ā£ How should requests be scheduled across replicas? Uniļ¬ed Solution: Back-pressure style scheduling High ThresholdLow Threshold
- 165. BlowFish: Technical Details ā£ How should partitions share cache on a server? ā£ How should partitions share cache across servers? Cache proportional to load, ā£ How should requests be scheduled across replicas? Uniļ¬ed Solution: Back-pressure style scheduling High ThresholdLow Threshold
- 166. BlowFish: Technical Details ā£ How should partitions share cache on a server? ā£ How should partitions share cache across servers? Cache proportional to load, ā£ How should requests be scheduled across replicas? Uniļ¬ed Solution: Back-pressure style scheduling without explicit coordination High ThresholdLow Threshold
- 167. BlowFish: Technical Details ā£ How should partitions share cache on a server? ā£ How should partitions share cache across servers? ā£ How should requests be scheduled across replicas? Uniļ¬ed Solution: Back-pressure style scheduling 1.5x higher throughput than Selective Replication, High ThresholdLow Threshold
- 168. BlowFish: Technical Details ā£ How should partitions share cache on a server? ā£ How should partitions share cache across servers? ā£ How should requests be scheduled across replicas? Uniļ¬ed Solution: Back-pressure style scheduling 1.5x higher throughput than Selective Replication, within 11% of maximum possible throughput High ThresholdLow Threshold
- 169. Succinct + BlowFish
- 170. ā£ Standalone system (prototyped & tested) Succinct + BlowFish
- 171. ā£ Standalone system (prototyped & tested) ā£ Spark Package: Succinct on Apache Spark Succinct + BlowFish
- 172. ā£ Standalone system (prototyped & tested) ā£ Spark Package: Succinct on Apache Spark ā£ As libraries ā£ C++, Java, Scala ā£ for ease of integration Succinct + BlowFish
- 174. Backup Slides
- 175. Array of Suļ¬xes (AoS) banana$ (Input)
- 176. Array of Suļ¬xes (AoS) banana$ banana$ anana$ nana$ ana$ na$ a$ $ Sufļ¬xes (Input)
- 177. Array of Suļ¬xes (AoS) banana$ banana$ anana$ nana$ ana$ na$ a$ $ Sufļ¬xes $ a$ ana$ anana$ banana$ na$ nana$ Array of Sufļ¬xes (AoS) lexicographicalorder (Input)
- 178. AoS to Input (AoS2Input) Array $ a$ ana$ anana$ banana$ na$ nana$ AoS 6 AoS2Input 5 3 1 0 4 2 b Input 0 1 2 3 4 5 6 a n a n a $
- 179. AoS to Input (AoS2Input) Array $ a$ ana$ anana$ banana$ na$ nana$ AoS 6 AoS2Input 5 3 1 0 4 2 b Input 0 1 2 3 4 5 6 a n a n a $ locations of sufļ¬xes (sufļ¬x array)
- 180. AoS to Input (AoS2Input) Array $ a$ ana$ anana$ banana$ na$ nana$ AoS 6 AoS2Input 5 3 1 0 4 2 b Input 0 1 2 3 4 5 6 a n a n a $ locations of sufļ¬xes (sufļ¬x array)
- 181. AoS to Input (AoS2Input) Array $ a$ ana$ anana$ banana$ na$ nana$ AoS 6 AoS2Input 5 3 1 0 4 2 b Input 0 1 2 3 4 5 6 a n a n a $ locations of sufļ¬xes (sufļ¬x array)
- 189. Next Pointer Array: Reducing AoS Size $ a$ ana$ anana$ banana$ na$ nana$ AoS 0 1 2 3 4 5 6 NPA
- 190. Next Pointer Array: Reducing AoS Size $ a$ ana$ anana$ banana$ na$ nana$ AoS 0 1 2 3 4 5 6 NPA 3
- 191. Next Pointer Array: Reducing AoS Size $ a$ ana$ anana$ banana$ na$ nana$ AoS 0 1 2 3 4 5 6 NPA 3
- 192. Next Pointer Array: Reducing AoS Size $ a$ ana$ anana$ banana$ na$ nana$ AoS 0 1 2 3 4 5 6 NPA 3 6
- 193. Next Pointer Array: Reducing AoS Size $ a$ ana$ anana$ banana$ na$ nana$ AoS 0 1 2 3 4 5 6 NPA 3 6
- 194. Next Pointer Array: Reducing AoS Size $ a$ ana$ anana$ banana$ na$ nana$ AoS 0 1 2 3 4 5 6 NPA 2 3 6
- 195. Next Pointer Array: Reducing AoS Size $ a$ ana$ anana$ banana$ na$ nana$ AoS 0 1 2 3 4 5 6 NPA 2 3 6
- 196. Next Pointer Array: Reducing AoS Size $ a$ ana$ anana$ banana$ na$ nana$ AoS 0 1 2 3 4 5 6 NPA 4 0 5 1 2 3 6
- 197. Next Pointer Array: Reducing AoS Size $ a$ ana$ anana$ banana$ na$ nana$ AoS 0 1 2 3 4 5 6 NPA 4 0 5 1 2 AoS NPA $0 1 2 3 4 5 6 a a a b n n 4 0 5 6 3 1 2 3 6
- 198. Next Pointer Array: Reducing AoS Size $ a$ ana$ anana$ banana$ na$ nana$ AoS 0 1 2 3 4 5 6 NPA 4 0 5 1 2 AoS NPA $0 1 2 3 4 5 6 a a a b n n 4 0 5 6 3 1 2 Store only the ļ¬rst character (entire sufļ¬x can be computed āon the ļ¬yā using Next Pointer Array (NPA)) 3 6
- 199. Next Pointer Array: Reducing AoS Size $ a$ ana$ anana$ banana$ na$ nana$ AoS 0 1 2 3 4 5 6 NPA 4 0 5 1 2 AoS NPA $0 1 2 3 4 5 6 a a a b n n 4 0 5 6 3 1 2 3 6
- 200. Next Pointer Array: Reducing AoS Size $ a$ ana$ anana$ banana$ na$ nana$ AoS 0 1 2 3 4 5 6 NPA 4 0 5 1 2 AoS NPA $0 1 2 3 4 5 6 a a a b n n 4 0 5 6 3 1 2 3 6 a
- 201. Next Pointer Array: Reducing AoS Size $ a$ ana$ anana$ banana$ na$ nana$ AoS 0 1 2 3 4 5 6 NPA 4 0 5 1 2 AoS NPA $0 1 2 3 4 5 6 a a a b n n 4 0 5 6 3 1 2 3 6 an
- 202. Next Pointer Array: Reducing AoS Size $ a$ ana$ anana$ banana$ na$ nana$ AoS 0 1 2 3 4 5 6 NPA 4 0 5 1 2 AoS NPA $0 1 2 3 4 5 6 a a a b n n 4 0 5 6 3 1 2 3 6 an
- 203. Next Pointer Array: Reducing AoS Size $ a$ ana$ anana$ banana$ na$ nana$ AoS 0 1 2 3 4 5 6 NPA 4 0 5 1 2 AoS NPA $0 1 2 3 4 5 6 a a a b n n 4 0 5 6 3 1 2 3 6 ana
- 204. Next Pointer Array: Reducing AoS Size $ a$ ana$ anana$ banana$ na$ nana$ AoS 0 1 2 3 4 5 6 NPA 4 0 5 1 2 AoS NPA $0 1 2 3 4 5 6 a a a b n n 4 0 5 6 3 1 2 3 6 ana
- 205. Next Pointer Array: Reducing AoS Size $ a$ ana$ anana$ banana$ na$ nana$ AoS 0 1 2 3 4 5 6 NPA 4 0 5 1 2 AoS NPA $0 1 2 3 4 5 6 a a a b n n 4 0 5 6 3 1 2 3 6 ana$
- 206. Next Pointer Array: Reducing AoS Size $ a$ ana$ anana$ banana$ na$ nana$ AoS 0 1 2 3 4 5 6 NPA 4 0 5 1 2 AoS NPA $0 1 2 3 4 5 6 a a a b n n 4 0 5 6 3 1 2 3 6 ana$
- 207. Next Pointer Array: Reducing AoS Size $ a$ ana$ anana$ banana$ na$ nana$ AoS 0 1 2 3 4 5 6 NPA 4 0 5 1 2 AoS NPA $0 1 2 3 4 5 6 a a a b n n 4 0 5 6 3 1 2 3 6
- 208. Next Pointer Array: Reducing AoS Size $ a$ ana$ anana$ banana$ na$ nana$ AoS 0 1 2 3 4 5 6 NPA 4 0 5 1 2 AoS NPA $ a b n 4 0 5 6 3 1 2 0 1 2 3 4 5 6 AoS NPA $0 1 2 3 4 5 6 a a a b n n 4 0 5 6 3 1 2 3 6
- 209. Reducing the size of AoS2Input 6 AoS2Input 5 0 2 4 NPA 0 5 6 3 1 2 0 1 2 3 4 5 6 3 1 4
- 210. Reducing the size of AoS2Input 6 AoS2Input 5 0 2 4 NPA 0 5 6 3 1 2 0 1 2 3 4 5 6 3 1 4
- 211. Reducing the size of AoS2Input 6 AoS2Input 5 0 2 4 NPA 0 5 6 3 1 2 0 1 2 3 4 5 6 3 1 4
- 212. Reducing the size of AoS2Input 6 AoS2Input 5 0 2 4 NPA 0 5 6 3 1 2 0 1 2 3 4 5 6 3 1 4 AoS2Input NPA 4 0 5 6 3 1 2 6 0 2 0 1 2 3 4 5 6 3
- 213. Reducing the size of AoS2Input 6 AoS2Input 5 0 2 4 NPA 0 5 6 3 1 2 0 1 2 3 4 5 6 3 1 4 AoS2Input NPA 4 0 5 6 3 1 2 6 0 2 0 1 2 3 4 5 6 3 Store only a few sampled values (unsampled values computed āon the ļ¬yā using NPA)
- 214. Reducing the size of AoS2Input 6 AoS2Input 5 0 2 4 NPA 0 5 6 3 1 2 0 1 2 3 4 5 6 3 1 4 AoS2Input NPA 4 0 5 6 3 1 2 6 0 2 0 1 2 3 4 5 6 3 Store only a few sampled values (unsampled values computed āon the ļ¬yā using NPA)
- 215. Reducing the size of AoS2Input 6 AoS2Input 5 0 2 4 NPA 0 5 6 3 1 2 0 1 2 3 4 5 6 3 1 4 AoS2Input NPA 4 0 5 6 3 1 2 6 0 2 0 1 2 3 4 5 6 3 Store only a few sampled values (unsampled values computed āon the ļ¬yā using NPA)
- 216. Compressing NPA Increasing sequence of integers (values for sufļ¬xes starting with same character) Can be compressed (E.g., using run-length encoding) $ a b n 4 0 5 6 3 1 2
- 217. Compressing NPA Increasing sequence of integers (values for sufļ¬xes starting with same character) Can be compressed (E.g., using run-length encoding) Succinct uses a 2-dimensional representation of NPA $ a b n 4 0 5 6 3 1 2
- 218. Compressing NPA Increasing sequence of integers (values for sufļ¬xes starting with same character) Can be compressed (E.g., using run-length encoding) Succinct uses a 2-dimensional representation of NPA $ a b n 4 0 5 6 3 1 2
- 219. Compressing NPA Increasing sequence of integers (values for sufļ¬xes starting with same character) Can be compressed (E.g., using run-length encoding) Succinct uses a 2-dimensional representation of NPA - better compressibility $ a b n 4 0 5 6 3 1 2
- 220. Compressing NPA Increasing sequence of integers (values for sufļ¬xes starting with same character) Can be compressed (E.g., using run-length encoding) Succinct uses a 2-dimensional representation of NPA - better compressibility - avoids binary search on AoS (lower latency) $ a b n 4 0 5 6 3 1 2
- 221. Compressing NPA Increasing sequence of integers (values for sufļ¬xes starting with same character) Can be compressed (E.g., using run-length encoding) Succinct uses a 2-dimensional representation of NPA - better compressibility - avoids binary search on AoS (lower latency) - enables wider range of queries (E.g., RegEx) $ a b n 4 0 5 6 3 1 2
- 222. Compressing NPA Increasing sequence of integers (values for sufļ¬xes starting with same character) Can be compressed (E.g., using run-length encoding) Succinct uses a 2-dimensional representation of NPA - better compressibility - avoids binary search on AoS (lower latency) - enables wider range of queries (E.g., RegEx) $ a b n 4 0 5 6 3 1 2
- 223. Compressing NPA Increasing sequence of integers (values for sufļ¬xes starting with same character) Can be compressed (E.g., using run-length encoding) Succinct uses a 2-dimensional representation of NPA - better compressibility - avoids binary search on AoS (lower latency) - enables wider range of queries (E.g., RegEx) See upcoming NSDI paper! $ a b n 4 0 5 6 3 1 2
- 224. Evaluation: Storage Footprint 10 node 150GB cluster
- 225. Evaluation: Storage Footprint 10 with in- h metadata; r-store with a bug also es the sys- ariable. For ss to the in- orm micro- single ma- failure sce- nd Cassan- exes. These nd wildcard rt wildcard ide slightly y, for Suc- valuate the tion. lti-attribute rgeKV from 75 150 225 DataSizethat FitsinMemory(GB) SmallKV LargeKV MongoDB Cassandra HyperDex Succinct RAM Figure 12: Succinct pushes more than 10Ć larger amount of data in memory compared to the next best system, while providing similar or stronger functionality. 10 node 150GB cluster
- 226. Evaluation: Storage Footprint Takeaway: Succinct can push >11x more data in memory 10 with in- h metadata; r-store with a bug also es the sys- ariable. For ss to the in- orm micro- single ma- failure sce- nd Cassan- exes. These nd wildcard rt wildcard ide slightly y, for Suc- valuate the tion. lti-attribute rgeKV from 75 150 225 DataSizethat FitsinMemory(GB) SmallKV LargeKV MongoDB Cassandra HyperDex Succinct RAM Figure 12: Succinct pushes more than 10Ć larger amount of data in memory compared to the next best system, while providing similar or stronger functionality. 10 node 150GB cluster
- 227. Evaluation: Throughput (95% GET + 5% PUT) 10 node 150GB cluster, uniform random access pattern
- 228. Evaluation: Throughput (95% GET + 5% PUT) 10 node 150GB cluster, uniform random access pattern
- 229. Evaluation: Throughput (95% GET + 5% PUT) Takeaway: Succinct achieves performance comparable to existing open source systems for queries on primary attributes 10 node 150GB cluster, uniform random access pattern
- 230. Evaluation: Throughput (95% SEARCH + 5% PUT) 10 node 150GB cluster, search queries with 1-10K occurrences
- 231. Evaluation: Throughput (95% SEARCH + 5% PUT) 10 node 150GB cluster, search queries with 1-10K occurrences
- 232. Evaluation: Throughput (95% SEARCH + 5% PUT) Takeaway: Succinct by pushing more data in faster storage provides performance similar to existing systems for 10-11x larger data sizes. 10 node 150GB cluster, search queries with 1-10K occurrences
- 233. Evaluation: RegEx Latency 40GB Wikipedia dataset, 5 commonly used RegEx queries Single EC2 node, 32 vCPUs, 244GB RAM
- 234. Evaluation: RegEx Latency 40GB Wikipedia dataset, 5 commonly used RegEx queries Single EC2 node, 32 vCPUs, 244GB RAM
- 235. Evaluation: RegEx Latency Takeaway: Succinct signiļ¬cantly speeds up RegEx queries even when all the data ļ¬ts in memory for all systems. 40GB Wikipedia dataset, 5 commonly used RegEx queries Single EC2 node, 32 vCPUs, 244GB RAM
- 236. Support for JSON
- 237. val ids1 = succinctJsonRDD.search("AMPLab") Search for JSON documents containing āAMPLabā Support for JSON
- 238. val ids2 = succinctJsonRDD.filter("city", "Berkeley") val ids1 = succinctJsonRDD.search("AMPLab") Filter JSON documents where the ācityā attribute has value āBerkeleyā Support for JSON
- 239. val jsonDoc = succinctJsonRDD.get(0) val ids2 = succinctJsonRDD.filter("city", "Berkeley") val ids1 = succinctJsonRDD.search("AMPLab") Get JSON document with id 0 Support for JSON
- 240. Layer Additions & Deletions
- 241. 9 12RATE = 8 3 14RATE = 4 15 0 8 5RATE = 2 Layer Additions & Deletions
- 242. 9 12RATE = 8 3 14RATE = 4 Layer Additions & Deletions Layer Deletions: simple
- 243. RATE = 2 9 12RATE = 8 3 14RATE = 4 Layer Additions & Deletions Layer Addition:
- 244. RATE = 2 9 12RATE = 8 3 14RATE = 4 Unsampled values already computed during query execution Layer Additions & Deletions Layer Addition:
- 245. RATE = 2 9 12RATE = 8 3 14RATE = 4 815 Unsampled values already computed during query execution Layer Additions & Deletions Layer Addition: Layers in LSA populated opportunistically!!
- 246. Spatial Skew
- 247. Spatial Skew Load distribution across partitions is heavily skewed
- 248. Object Load 1 Compressed Wasted Cache! Spatial Skew Load distribution across partitions is heavily skewed #Replicas Ī± Load Selective Replication
- 249. Spatial Skew Load distribution across partitions is heavily skewed #Replicas Ī± Load Selective Replication BlowFish Fractionally change storage just enough to meet load 1 Compressed Uncompressed 10 Object Load
- 250. Spatial Skew Load distribution across partitions is heavily skewed #Replicas Ī± Load Selective Replication BlowFish Fractionally change storage just enough to meet load 1.5x higher throughput than Selective Replication, 1 Compressed Uncompressed 10 Object Load
- 251. Spatial Skew Load distribution across partitions is heavily skewed #Replicas Ī± Load Selective Replication BlowFish Fractionally change storage just enough to meet load 1.5x higher throughput than Selective Replication, within 10% of optimal 1 Compressed Uncompressed 10 Object Load
- 252. Changes in Spatial Skew
- 253. Changes in Spatial Skew Study on Facebook Warehouse Cluster [HotStorageā13]
- 254. Changes in Spatial Skew Transient failures ā 90% of failuresStudy on Facebook Warehouse Cluster [HotStorageā13]
- 255. Changes in Spatial Skew Transient failures ā 90% of failures Replica creation delayed by 15 mins Study on Facebook Warehouse Cluster [HotStorageā13]
- 256. Changes in Spatial Skew Transient failures ā 90% of failures Replica creation delayed by 15 mins Study on Facebook Warehouse Cluster [HotStorageā13] Leads to variation in load over time
- 257. Changes in Spatial Skew Transient failures ā 90% of failures Replica creation delayed by 15 mins Replica#1 Replica#2 Replica#3 Data Partitions Request Queues Study on Facebook Warehouse Cluster [HotStorageā13] Leads to variation in load over time
- 258. Changes in Spatial Skew Transient failures ā 90% of failures Replica creation delayed by 15 mins Replica#1 Replica#2 Replica#3 Data Partitions Request Queues Study on Facebook Warehouse Cluster [HotStorageā13] Leads to variation in load over time
- 259. Changes in Spatial Skew Transient failures ā 90% of failures Replica creation delayed by 15 mins Replica#1 Replica#2 Replica#3 Data Partitions Request Queues Study on Facebook Warehouse Cluster [HotStorageā13] Leads to variation in load over time
- 260. Changes in Spatial Skew Replica#1 Replica#2 Replica#3
- 261. Changes in Spatial Skew Replica#1 Replica#2 Replica#3
- 262. Changes in Spatial SkewOperations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Replica#1 Replica#2 Replica#3
- 263. Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Changes in Spatial Skew Load Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Replica#1 Replica#2 Replica#3
- 264. Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Changes in Spatial Skew Load Throughput Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Replica#1 Replica#2 Replica#3
- 265. Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Changes in Spatial Skew Load Throughput Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Replica#1 Replica#2 Replica#3
- 266. Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Changes in Spatial Skew Load Throughput Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 RequestQueueSize 0K 10K 20K 30K 40K 50K Time (mins) 0 30 60 90 120 Replica#1 Replica#2 Replica#3
- 267. Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 RequestQueueSize 0K 10K 20K 30K 40K 50K Time (mins) 0 30 60 90 120 Changes in Spatial Skew Load Throughput Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 RequestQueueSize 0K 10K 20K 30K 40K 50K Time (mins) 0 30 60 90 120 Replica#1 Replica#2 Replica#3
- 268. Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 RequestQueueSize 0K 10K 20K 30K 40K 50K Time (mins) 0 30 60 90 120 RequestQueueSize 0K 10K 20K 30K 40K 50K Time (mins) 0 30 60 90 120 Changes in Spatial Skew Load Throughput Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 RequestQueueSize 0K 10K 20K 30K 40K 50K Time (mins) 0 30 60 90 120 Replica#1 Replica#2 Replica#3
- 269. Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 RequestQueueSize 0K 10K 20K 30K 40K 50K Time (mins) 0 30 60 90 120 RequestQueueSize 0K 10K 20K 30K 40K 50K Time (mins) 0 30 60 90 120 Changes in Spatial Skew Load Throughput Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 RequestQueueSize 0K 10K 20K 30K 40K 50K Time (mins) 0 30 60 90 120 Replica#1 Replica#2 Replica#3
- 270. Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 RequestQueueSize 0K 10K 20K 30K 40K 50K Time (mins) 0 30 60 90 120 RequestQueueSize 0K 10K 20K 30K 40K 50K Time (mins) 0 30 60 90 120 Changes in Spatial Skew Load Throughput Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 RequestQueueSize 0K 10K 20K 30K 40K 50K Time (mins) 0 30 60 90 120 Replica#1 Replica#2 Replica#3
- 271. Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 RequestQueueSize 0K 10K 20K 30K 40K 50K Time (mins) 0 30 60 90 120 RequestQueueSize 0K 10K 20K 30K 40K 50K Time (mins) 0 30 60 90 120 RequestQueueSize 0K 10K 20K 30K 40K 50K Time (mins) 0 30 60 90 120 Changes in Spatial Skew Load Throughput Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 RequestQueueSize 0K 10K 20K 30K 40K 50K Time (mins) 0 30 60 90 120 Replica#1 Replica#2 Replica#3
- 272. Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 RequestQueueSize 0K 10K 20K 30K 40K 50K Time (mins) 0 30 60 90 120 RequestQueueSize 0K 10K 20K 30K 40K 50K Time (mins) 0 30 60 90 120 RequestQueueSize 0K 10K 20K 30K 40K 50K Time (mins) 0 30 60 90 120 Changes in Spatial Skew Load Throughput Operations/second 0 600 1200 1800 2400 3000 Time (mins) 0 30 60 90 120 RequestQueueSize 0K 10K 20K 30K 40K 50K Time (mins) 0 30 60 90 120 Adapts to 3x higher load in < 5 mins Replica#1 Replica#2 Replica#3