Letgo Data Platform: A global overview
•
8 j'aime•971 vues
How to develop a Big Data platform around Spark.
Recommandé
Recommandé
Contenu connexe
Tendances
Tendances (9)
Similaire à Letgo Data Platform: A global overview
Similaire à Letgo Data Platform: A global overview (20)
Dernier
Dernier (20)
Letgo Data Platform: A global overview
- 1. 1 Letgo Data Platform A global overview
- 3. 3 LETGO DATA PLATFORM IN NUMBERS 500GB 530+ Data daily Event Types Events Processed Daily 565M 180TB Storage (S3) 8K Events Processed per Second < 1sec NRT Processing Time
- 7. 7 MOVING TO EVENTS Domain EventsTracking Events
- 8. 8 MOVING TO EVENTS “A Domain Event captures the memory of something interesting (external stimulus) which affects the domain” Martin Fowler Domain EventsTracking Events
- 9. 9 GO TO EVENTS { "data":{ "id":"1c4da9f0-605e-4708-a8e7-0f2c97dff16e", "type":"message_sent", "attributes":{ "id":"e0abcd6c-c027-489f-95bf-24796e421e8b", "conversation_id":"47596e76-0fb4-4155-9aeb-6a5ba14c9cef", "product_id":"a0ef64a5-0a4d-48b8-9124-dd57371128f5", "from_talker_id":"5mCK6K8VCc", "to_talker_ids":[ "def4442e-6df5-4385-938a-8180ddfb6c5e" ], "message_id":"e0abcd6c-c027-489f-95bf-24796e421e8b", "type":"offer", "message":"Is this item still available?", "belongs_to_spam_conversation":false, "sent_at":1509235499995 } }, "meta":{ "created_at":1509235500012, ... }
- 10. 10 THE PATH WE CHOSE
- 14. 14 DATA INGESTION: Kafka Connect Kafka Connect: Connectors
- 17. 17 STORAGE We want to store all events coming from Kafka to S3
- 18. 18 STORAGE
- 19. 19
- 23. 23 Late events
- 25. 25 1 2 3 4 1 5 6 Read batch of events from Kafka. Write each event to Cassandra. Write dirty “hours” to compact topic: Key=(event_type, hour). Read dirty “hours” topic. Read all events with dirty hours. Store in S3 1 2 3 4 1 5 6 STORAGE: Late events Dirty Buckets
- 26. 26 STORAGE: Late events Some time later: Count events in Cassandra and S3 and check inconsistencies.
- 27. 27 S3 Big data problems
- 28. 28 1. Eventual consistency 2. Very slow renames: Rename = copy + delete. Some S3 Big Data problems: STORAGE: S3 Big data problems Why?
- 29. 29 1. Eventual consistency STORAGE: S3 Big data problems
- 30. 30 Available in: S3GUARD Solution STORAGE: S3 Big data problems
- 31. 31 ¿Job freeze? 2. Slow renames STORAGE: S3 Big data problems
- 32. 32 3.1 • New S3A Committers Solution STORAGE: S3 Big data problems
- 33. 33 • Committers Architecture • S3A Committers • Spark and S3 with Ryan Blue: SlideShare, YouTube STORAGE: S3 Big data problems
- 34. 34 STORAGE: S3 Big data problems Surprise!
- 37. 37 STREAMING: REAL-TIME USER SEGMENTATION Stream Journal User bucket’s changed 1 2
- 39. 39 STREAMING: REAL-TIME USER STATISTICS What we do: • Sessions (Last session) • Conversation (Last conversation created, counts, …) • Products (Last product approved, counts, ...) • User classification (Last bucket changed, counts, ...) Tool:
- 41. 41 STREAMING: REAL-TIME PATTERN DETECTION Is it still available? Is the price negotiable? What condition is it in? I offer you….$ Could we meet at……?
- 42. 42 STREAMING: REAL-TIME PATTERN DETECTION Event A + Event B = Complex Event C Event A + NOTHING in 2 hours = Complex Event D Some common use cases: • Fraud detection, ¿Scammers detection? • Real time recommendations
- 43. 43 GO TO EVENTS: Tips INGEST Data Ingestion Storage PROCESSING Stream Batch DISCOVER Query Data exploitation Orchestration
- 45. 45 BATCH: Geodata enrichment { "data": { "id": "105dg3272-8e5f-426f-bca0- 704e98552961", "type": "some_event", "attributes": { "latitude": 39.740028705904, "longitude": -104.97341156236 } }, "meta": { "created_at": 1522886400036 } }
- 46. 46 POINT (77.3548351 28.6973627) Coordinates: Longitude and latitude. What we know: BATCH: Geodata enrichment
- 47. 47 Where is this point? City, State, zip code, DMA. BATCH: Geodata enrichment
- 48. 48 POLYGON((-114.816294 32.508038,-114.814321 32.509023,-114.810159 32.508383,-114.807726 32.508726,-114.805239 32.509985... We represent in this way: States, Cities, Zip codes, DMAs What we wanted to know: BATCH: Geodata enrichment
- 49. 49 How we do it: • We load Cities, States, Zip codes, DMAs… polygons from Well-known text (WKT) • Create indexes using JTS Topology Suite • Custom Spark SQL UDF. Tool: BATCH: Geodata enrichment SELECT geodata.dma_name, geodata.dma_number AS dma_number, geodata.city AS city, geodata.state AS state , geodata.zip_code AS zip_code FROM ( SELECT geodata(longitude, latitude) AS geodata FROM …. )
- 51. 51 QUERYING DATA
- 52. 52 QUERYING DATA
- 54. 54 WHY WE CHOSE SPARK THRIFT SERVER?
- 55. 55 QUERYING DATA: WHY WE CHOSE SPARK THRIFT SERVER?
- 56. 56 QUERYING DATA: WHY WE CHOSE SPARK THRIFT SERVER?
- 57. 57 QUERYING DATA: WHY WE CHOSE SPARK THRIFT SERVER? Thrift Server
- 58. 58 QUERYING DATA
- 59. 59 QUERYING DATA CREATE EXTERNAL TABLE IF NOT EXISTS database_name.table_name( some_column STRING..., dt DATE ) PARTITIONED BY (`dt`) USING PARQUET LOCATION 's3a://bucket-name/database_name/table_name' CREATE TABLE IF NOT EXISTS database_name.table_name( some_column STRING, ... dt DATE ) USING json PARTITIONED BY (`dt`) CREATE TABLE IF NOT EXISTS database_name.table_name using com.databricks.spark.redshift options ( dbtable 'schema.redshift_table_name', tempdir 's3a://redshift-temp/', url 'jdbc:redshift://xxxx.redshift.amazonaws.com:5439/letgo?user=xxx&password=xx x', forward_spark_s3_credentials 'true' ) CREATE TEMPORARY VIEW table_name USING org.apache.spark.sql.cassandra OPTIONS ( table "table_name", keyspace "keyspace_name" )
- 60. 60 QUERYING DATA CREATE TABLE … USING [parquet,json,csv…] CREATE TABLE … STORED AS… VS
- 61. 61 QUERYING DATA CREATE TABLE … STORED AS… VS 70% Higher performance! CREATE TABLE … USING [parquet,json,csv…]
- 63. 63 QUERYING DATA: Batches with SQL CREATE EXTERNAL TABLE IF NOT EXISTS database.some_name( user_id STRING, column_b STRING, ... ) USING PARQUET PARTITIONED BY (`dt` STRING) LOCATION 's3a://example/some_table' 1. Creating the table
- 64. 64 QUERYING DATA: Batches with SQL INSERT OVERWRITE TABLE database.some_name PARTITION(dt) SELECT user_id, column_b, dt FROM other_table ... 2. Inserting data
- 65. 65 QUERYING DATA: Batches with SQL
- 66. 66 QUERYING DATA: Batches with SQL 200 files because default value of “spark.sql.shuffle.partition”
- 67. 67 QUERYING DATA: Batches with SQL INSERT OVERWRITE TABLE database.some_name PARTITION(dt) SELECT user_id, column_b, dt FROM other_table ... Solution
- 68. 68 QUERYING DATA: Batches with SQL DISTRIBUTE BY (dt): Only one file not Sorted CLUSTERED BY (dt, user_id, column_b): Multiple files DISTRIBUTE BY (dt) SORT BY (user_id, column_b): Only one file sorted by user_id, column_b. Good for joins using this properties.
- 69. 69 QUERYING DATA: Batches with SQL INSERT OVERWRITE TABLE database.some_name PARTITION(dt) SELECT user_id, column_b, dt FROM other_table ... DISTRIBUTE BY (dt) SORT BY (user_id)
- 70. 70 QUERYING DATA: Batches with SQL • Hive Bucketing in Apache Spark: SlideShare, YouTube.
- 75. 75 DATA SCIENTISTS (WINTER IS COMING)
- 76. 76 DATA EXPLOITATION: Data Scientists
- 77. 77 DATA EXPLOITATION: Data Scientists
- 78. 78
- 79. 79 DATA EXPLOITATION: Data Scientists Too many small files!
- 80. 80 DATA EXPLOITATION: Data Scientists Huge Query!
- 81. 81 DATA EXPLOITATION: Data Scientists Too much shuffle!
- 82. 82 DATA EXPLOITATION: Data Scientists
- 83. 83 Some data scientist tips Listen carefully or you will die!
- 84. 84 Transversal workshops: BI, Data Scientist, Data engineer. Create a wiki with: Guidelines, tips, links... Learn from your mistakes! DATA EXPLOITATION: Data Scientists Well done!
- 86. 86 ORCHESTRATION • It’s an open source project writen in Python by Airbnb. • Easy to Schedule Jobs, write new tasks and monitoring them.
- 87. 87 ORCHESTRATION
- 89. 89 TIPS
- 91. 91 TIPS: Bigger is better 20 slots Total used: 3x6x4=72 Total wasted: 2x4=8 Total used: 3x26=78 Total wasted: 2 80 slots We prefer big Hadoop instances. We use m4.16xlarge instances. We are migrating to m5.24xlarge instances.
- 94. 94 Be polite
- 95. 95 TIPS: Be polite spark.executor.memory = 4G spark.executor.cores = 3 spark.driver.memory = 4G spark.driver.cores = 2 ## dynamicAllocation spark.shuffle.service.enabled=true spark.dynamicAllocation.enabled=true spark.executor.instances=1 spark.dynamicAllocation.minExecutors=1 spark.dynamicAllocation.initialExecutors=3 spark.dynamicAllocation.maxExecutors=20 spark.dynamicAllocation.cachedExecutorIdleTimeout=60s Others are using the same cluster, don´t over reclaim resources.
- 96. 96 But not too polite
- 97. 97 TIPS: But not too polite
- 99. 99 TIPS: SPARK METRICS What we did: • Custom InfluxDB Spark metrics sink • Custom Spark metrics using Dropwizard • Grafana Dashbords and alerts
- 101. 101 TIPS: Mistakes are good
- 102. 102 Do you want to join us?