ACID ORC, Iceberg, and Delta Lake—An Overview of Table Formats for Large Scale Storage and Analy

The reality of most large scale data deployments includes storage decoupled from computation, pipelines operating directly on files and metadata services with no locking mechanisms or transaction tracking. For this reason attempts at achieving transactional behavior, snapshot isolation, safe schema evolution or performant support for CRUD operations has always been marred with tradeoffs.

This talk will focus on technical aspects, practical capabilities and the potential future of three table formats that have emerged in recent years as solutions to the issues mentioned above – ACID ORC (in Hive 3.x), Iceberg and Delta Lake. To provide a richer context, a comparison between traditional databases and big data tools as well as an overview of the reasons for the current state of affairs will be included.

After the talk, the audience is expected to have a clear understanding of the current development trends in large scale table formats, on the conceptual and practical level. This should allow the attendees to make better informed assessments about which approaches to data warehousing, metadata management and data pipelining they should adapt in their organizations.


1. ACID ORC, Iceberg and Delta Lake an overview of table formats for large scale storage and analytics Michal Gancarski wssbck 17-10-2019

2.TABLE OF CONTENTS All Is Not Well In The Land Of Big Data There Is Hope, However This Is How We Do It Moving Forward 2

3.All Is Not Well In The Land Of Big Data 3

4. ACID Properties 4

5. Single Node Database 5

6. Distributed Database 6

7. Distributed Data Infrastructure 7

8. Lost ACID 8

9.There Is Hope, However 9

10. A Table Format? 10

11. ACID ORC 11

12. ACID ORC CREATE TABLE d_manufacturers (id int, name string) PARTITIONED BY (country string) STORED AS ORC TBLPROPERTIES ('transactional'='true'); ./d_manufacturers/country=de/base_00000002/ -- bucket_00000 -- bucket_00001 ./d_manufacturers/country=de/delta_0000003_0000003_0000/ -- bucket_00000 -- bucket_00001 ./d_manufacturers/country=de/delta_0000004_0000004_0000/ -- bucket_00001 ./d_manufacturers/country=de/delete_delta_0000004_0000004_0000/ -- bucket_00001 12

13. ACID ORC + ❖ Native compatibility with Hive ❖ Fast updates / upserts (no file rewrite) ❖ Hive 2.x ACID ORC tables can be converted to Hive 3.x ACID ORC tables ❖ Commercial Support (Cloudera) ❖ Limited support for Spark (being worked on by Qubole) - ❖ Slow listing and metadata discovery ❖ Potentially slower read due to ad-hoc compaction ❖ ORC only ❖ Mandatory S3Guard or EMR with consistent view enabled 13

14. Apache Iceberg 14

15. Apache Iceberg val df = .format("iceberg") .load("s3://datalake/d_manufacturers") 15

16. Apache Iceberg ❖ Parquet, Avro, ORC supported as file formats ❖ Robust schema and partitioning changes + ❖ Fast query planning ❖ Presto connector ❖ Time travel with snapshot id listing ❖ No dependency on Spark public List<Snapshot> snapshots() { return snapshots; - } ❖ Spark support ❖ Sparse documentation ❖ No commercial support ❖ Not as mature as other formats 16

17. Delta Lake 17

18. Delta Lake val df = .format("delta") .load("s3://datalake/d_manufacturers") CONVERT TO DELTA parquet.`s3://datalake/d_manufacturers` ./d_manufacturers/_delta_log/ -- 000000.json -- ... -- 000010.checkpoint.parquet -- _latest_checkpoint ./d_manufacturers/country=de/ -- file_1.parquet -- file_2.parquet ./d_manufacturers/country=fr/ -- file_3.parquet 18

19. Delta Lake ❖ Great integration with Spark, including Structured + Streaming ❖ Merge syntax in Spark SQL ❖ Time travel ❖ Comprehensive, well written documentation ❖ Fast development backed by a commercial entity ❖ VACUUM + OPTIMIZE ❖ Incoming Presto reader (Starburst) - ❖ ❖ Parquet only Multicluster writes outside of Databricks only on HDFS 19

20.This Is How We Do It 20

21. Delta Lake @Zalando 21

22.Moving Forward 22

23. The Future is Bright 23

24. Further Reading ACID ORC Iceberg Delta Lake 24

25. Further Reading Engine Support S3 Consistency Other 25

26.ACID ORC, Iceberg and Delta Lake Michal Gancarski wssbck 26