1. What to do with Scientific Data? Michael Stonebraker

2. Outline Science data -- what it looks like Software options - File system - RDBMS - SciDB

3. O(100) petabytes

5. LSST Data Raw imagery - 2D arrays of telescope readings Cooked into observations - Image intensity algorithm (data clustering) - Spatial data Further cooked into trajectories - Similarity query - Constrained by maximum distance

6. Example LSST Queries Re-cook raw imagery with my algorithm Find all observations in a spatial region Find all trajectories that intersect a cylinder in time

7. Chemical Plant Data Plant is a directed graph of plumbing Sensors at various places - 1/sec observations Directed graph of time series Plant runs near the edge to optimize output And fails every once in a while -- down for a week

8. Chemical Plant Data Record all data {(time, sensor-1, …, sensor-5000)} Look for interesting events - i.e. sensor values out of whack Cluster events near each other in 5000 dimension space Goal is to identify near-failure modes

9. Snow Cover in the Sierras Snow Cover in the Sierras

10. Earth Science Data Raw imagery – arrays cooked into composite images based on multiple passes - Cloud free cells baked into snow cover, …

11. General Model General Model sensors Cooking Algorithm(s) (pipeline) Derived data

12. Traditional Wisdom (1) Cooking pipeline in hard code or custom hardware All data in some file system

13. Problems Cant find anything Metadata often not recorded - Sensor parameters - Cooking parameters Cant easily share anything Can easily recook anything Everything is custom code

14. Traditional Wisdom (2) Cooking pipeline outside DBMS Derived data loaded into DBMS for subsequent querying

15. Problems with This Approach Most of the issues remain Better, but stay tuned

16. My Preference Load the raw data into a DBMS Cooking pipeline is a collection of user-defined functions (DBMS extensions) Activated by triggers or a workflow management system ALL data captured in a common system!!!!

17. What DBMS to Use? RDBMS (e.g. Oracle) - Pretty hopeless on raw data Simulating arrays on top of tables likely to cost a factor of 10-100 - Not pretty on time series data Find me a sensor reading whose average value over the last 3 days is within 1% of the average value adjoining 5 sensors

18. RDBMS Summary Wrong data model - Arrays not tables Wrong operations - Regrid not join Missing features - Versions, no-overwrite, provenance, support for uncertain data, …

19. But your mileage may vary SQL Server working well for Sloan Skyserver database See paper in CIDR 2009 by Jose Blakeley

20. How to Do Analytics (e.g. clustering) Suck out the data Convert to array format Pass to Matlab, SAS, R Compute Return answer to DBMS

21. Bad News Painful Slow Many analysis platforms are not parallel Many are main memory only

22. My Proposal - SciDB Build a commercial-quality array DBMS from the ground up with integrated analytics - Open source!!!!

23. Data Model Nested multi-dimensional arrays - Cells can be tuples or other arrays Time is an automatically supported extra dimension Ragged arrays allow each row/column to have a different dimensionality Support for multiple flavors of null - Array cells can be EMPTY - User-definable, context- sensitive treatment

24. Array-oriented storage manager Optimized for both dense and sparse array data - Different data storage, compression, and access Arrays are chunked (in multiple dimensions) Chunks are partitioned across a collection of nodes Chunks have overlap to support neighborhood operations Replication provides efficiency and back-up No overwrite

25. Doesnt require JDBC, ODBC AQL an extension of SQL Also supports UDFs Java, C++, whatever… Architecture Shared nothing cluster - 10s–1000s of nodes - Commodity hardware - TCP/IP between nodes - Linear scale-up Each node has a processor and storage Queries refer to arrays as if not distributed Query planner optimizes queries for efficient data access & processing Query plan runs on a nodes local executor&storage manager Runtime supervisor coordinates execution

26. SciDB DDL CREATE ARRAY Test_Array < A: integer NULLS, B: double, C: USER_DEFINED_TYPE > [I=0:99999,1000, 10, J=0:99999,1000, 10 ] PARTITION OVER ( Node1, Node2, Node3 ) USING block_cyclic(); chunk size 1000 overlap 10 attribute names A, B, C index names I, J

27. Array Query Language (AQL) Array data management - e.g. filter, aggregate, join, etc. Statistical & linear algebra operations - multiply, QR factorization, etc. - parallel, disk-oriented User-defined operators (Postgres-style) Interface to external stat packages (e.g. R)

28. Array Query Language (AQL) SELECT Geo-Mean ( T.B ) FROM Test_Array T WHERE T.I BETWEEN :C1 AND :C2 AND T.J BETWEEN :C3 AND :C4 AND T.A = 10 GROUP BY T.I; User-defined aggregate on an attribute B in array T Subsample Filter Group-by So far as SELECT / FROM / WHERE / GROUP BY queries are concerned, there is little logical difference between AQL and SQL

29. Matrix Multiply Algorithm sensitive to distribution (range, BC) For range, the smaller of the two arrays is replicated at all nodes - Scatter-gather Each node does its core of the bigger array with the replicated smaller one Produces a distributed answer CREATE ARRAY TS_Data [ I=0:99999,1000,0, J=0:3999,100,0 ] multiply (project (TS_Data, B1), transpose(project (TS_Data, B1))) 10K x 4K Project on one attribute

30. Other Features Science Guys Want (These could be in an RDBMS, but arent) Uncertainty - Data has error bars - Which must be carried along in the computation interval arithmentic

31. Other Features Time travel - Dont fix errors by overwrite - i.e. keep all the data – in the extra time dimension Named versions - Recalibration usually handled this way

32. Other Features Provenance (lineage) - What calibration generated the data - What was the cooking algorithm - In general - repeatabiltiy of derviced data

33. SciDB status www.scidb.org Global development team (17 people) including many volunteers Support/distribution by a VC-backed company - Paradigm4 in Waltham 0.75 on website now - Have at it!!! 1.0 release will be on website within a month - Very functional, but a little pokey Red Queen will be out in October - Blazing performance - Probably 5X current system

34. Performance of SciDB 1.0 Performance on stat operations - Comparable to R, degrades in a similar way when data does not fit in main memory - But multi-node parallelism provided -- much better scalability Performance on SQL - Comparable to or better than Postgres, especially on multi-attribute restrictions - And multi-node parallelism provided And we can do stuff they cant….