Massive Data Processing in Adobe Using Delta Lake
•
0 likes•779 views
At Adobe Experience Platform, we ingest TBs of data every day and manage PBs of data for our customers as part of the Unified Profile Offering. At the heart of this is a bunch of complex ingestion of a mix of normalized and denormalized data with various linkage scenarios power by a central Identity Linking Graph. This helps power various marketing scenarios that are activated in multiple platforms and channels like email, advertisements etc. We will go over how we built a cost effective and scalable data pipeline using Apache Spark and Delta Lake and share our experiences. What are we storing? Multi Source – Multi Channel Problem Data Representation and Nested Schema Evolution Performance Trade Offs with Various formats Go over anti-patterns used (String FTW) Data Manipulation using UDFs Writer Worries and How to Wipe them Away Staging Tables FTW Datalake Replication Lag Tracking Performance Time!
Report
Share
More Related Content
What's hot
What's hot (20)
Similar to Massive Data Processing in Adobe Using Delta Lake
Similar to Massive Data Processing in Adobe Using Delta Lake (20)
More from Databricks
More from Databricks (20)
Recently uploaded
Recently uploaded (20)
Massive Data Processing in Adobe Using Delta Lake
- 1. Massive Data Processing in Adobe using Delta Lake Yeshwanth Vijayakumar Sr. Engineering Manager/Architect @ Adobe
- 2. Agenda § Introduction § What are we storing? § Data Representation and Nested Schema Evolution § Writer Worries and How to Wipe them Away § Staging Tables FTW § Datalake Replication Lag Tracking § Performance Time!
- 3. Unified Profile Data Ingestion Unified Profile Experience Data Model Adobe Campaign AEM Adobe Analytics Adobe AdCloud Change Feed Streaming Stats Generation Single Tenant Multi Tenant
- 5. Data Layout At a Glance An Idea about how the graph linkages are stored primaryId relatedIds field1 field2 field1000 123 123 a b c 456 456 d e f 123 123 d e l 789 789,101 x y z 101 789,101 x u p Conditions • primaryId does not change • relatedIds can change
- 6. New Record comes in Indicating a new linkage, causing a change in graph membership primaryId relatedId field1 field2 field1000 103 103,789,101 q w r 789 103,789,101 x y z 101 103,789,101 x y z primaryId relatedId field1 field2 field1000 103 103,789,101 q w r New Record comes in linking 103 with 789 and 101 Causes a cascading change in rows of 789 and 101
- 7. Main Access Pattern Query 1 Query 2 Query 3 Query 1000 Multiple Queries over 1 consolidated row
- 8. Complexities? • Nested Fields • a.b.c.d[*].e nested hairiness! • Arrays! • MapType • Every Tenant has a different Schema! • Schema evolves constantly • Fields can get deleted, updated. • Multiple Sources • Streaming • Batch
- 9. Scale? • Tenants have 10+ Billions of rows • PBs of data • Million RPS peak across the system • Triggers multiple downstream applications • Segmentation • Activation
- 10. What is DeltaLake? From delta.io : Delta Lake is an open-source project that enables building a Lakehouse architecture on top of existing storage systems such as S3, ADLS, GCS, and HDFS. ACID Transactions Time Travel (data versioning) Uses Parquet Underneath Schema Enforcement and Schema Evolution Audit History Updates and Deletes Support Key Features
- 11. Delta lake in Practice UPSERT SQL Compatible
- 12. Writer Worries and How to Wipe them Away • Concurrency Conflicts • Column size • When individual column data exceeds 2GB, we see degradation in writes or OOM • Update frequency • Too frequent updates cause underlying filestore metadata issues. • This is because every transacation on an individual parquet causes CoW, • More updates => more rewrites on HDFS • Too Many small files !!!
- 13. CDC (existing) Batch Ingestion / Streaming Ingestion / API based Ingest Mutation Apps CosmosDB CDC 1. Send Request to Cosmos 2.Ack 3.Emit CDC Consumed by • Stats • Edge • etc
- 14. Dataflow with DeltaLake primary Id relatedId field 1 field2 field1000 103 103,789,101 q w r 789 103,789,101 x y z 101 103,789,101 x y z Cosmos DB primaryId relatedId field1 field1000 103 103,789,101 q r primaryId relatedId jsonString 103 103,789,101 <jsonStr> 789 103,789,101 <jsonStr> 101 103,789,101 <jsonStr> Staging Table Change Feed CDC Raw Table (per tenant) Check for Work every X minutes UPSERT/DELETE into Raw Table Fetch Records to process APPEND only! CDC Dumper Backfill Long Running Streaming Application Processor Partitioned by tenant and 15 min time intervals TenantLock in Redis
- 15. Staging Tables FTW Fan-In pattern vs Fan-out • Multiple Source Writers Issue Solved • By centralizing all reads from CDC, since ALL writes generate a CDC • Staging Table in APPEND ONLY mode • No conflicts while writing to it • Filter out. Bad data > thresholds before making it to Raw Table • Batch Writes by reading larger blocks of data from Staging Table • Since it acts time aware message buffer
- 16. Staging Table Logical View ProgressMap
- 17. Why choose JSON String format? § We are doing a lazy Schema on-read approach. ▪ Yes. this is an anti-pattern. § Nested Schema Evolution was not supported on update in delta in 2020 ▪ Supported with latest version § We want to apply conflict resolution before upsert-ing ▪ Eg. resolveAndMerge(newData, oldData) ▪ UDF’s are strict on types, with the plethora of difference schemas , it is crazy to manage UDF per org in Multi tenant fashion ▪ Now we just have simple JSON merge udfs ▪ We use json-iter which is very efficient in loading partial bits of json and in manipulating them. § Don’t you lose predicate pushdown? ▪ We have pulled out all main push-down filters to individual columns ▪ Eg. timestamp, recordType, id, etc. ▪ Profile workloads are mainly scan based since we can run 1000’s of queries at a single time. ▪ Reading the whole JSON string from datalake is much faster and cheaper than reading from Cosmos for 20% of all fields.
- 18. Schema On Read is more future safe approach for raw data § Wrangling Spark Structs is not user friendly § JSON schema is messy ▪ Crazy nesting ▪ Add maps to the equation, just the schema will be in MBs § Schema on Read using Json-iter means we can read what we need on a row by row basis § Materialized Views WILL have structs!
- 19. Partition Scheme of Raw records • RawRecords Delta Table • recordType • sourceId • timestamp (key-value records will use DEFAULT value) z-order on primaryId z-order - Colocate column information in the same set of files using locality-preserving space-filling curves
- 21. Replication Lag – 2 types • CDC Lag from Kafka • Tells us how much more work we need to do to catch up to write to Staging Table • How we track Lag on a per tenant basis • We track Max(TimeStamp) in CDC per org • We track Max(TSKEY) processed in Processor • Difference gives us rough lag of replication
- 22. Merge/UPSERT Performance Action: UPSERT CDC stage into fragment Time Taken 170 K CDC Records – Maps to 100k Rows in Raw Table 15 seconds 1.7 Million CDC Records – Maps to 1 Million Rows in Raw Table 61 seconds Live Traffic Usecase: How long does it take X CDC messages to get upserted into Raw Table
- 23. Job Performance Time! Hot Store (NoSQL Store) Delta Lake Size of Data 1 TB 64 GB Number of Partitions 80 189 Job Cores Used 112 112 Job Runtime 3 hours 25 mins
- 24. TakeAways • Scan IO speed from datalake >>> Read from Hot Store • Reasonably fast eventually consistent replication within minutes • More partitions means better Spark executor core utilization • Potential to aggressively TTL data in hot store • More downstream materialization !!! • Incremental Computation Framework thanks to Staging tables!
- 25. Feedback Your feedback is important to us. Don’t forget to rate and review the sessions.