Indexing OLAP Data Sunita Sarawagi

1
Indexing OLAP DataSunita Sarawagi

• Monowar Hossain
• York University

2
Agenda

• Requirements on Indexing methods
• Existing indexing methods
• Optimization of R-Tree for OLAP data
• R-Tree VS Bit-mapped Indices
• Conclusion

3
Requirements on Indexing methods

• Symmetric partial match queries
• Continuous e.g. time between Jan to July 94
• Discontinuous e.g. first month of each year
• Indexing at multiple levels of aggregation
• Pre-computation group-bys
• Indexing summary data
• Handing multiple traversal orders
• Efficient batch update
• Handling sparse data efficiently

4
Existing methods

• Multidimensional array-based methods
• Works efficiently when data is dense
• Essbases schema
• E.G. four dimensional cube product and store
  (sparse), time and scenarios ( dense)
• B-tree on Product and Store
• Two-dimensional array on time and scenarios
• Evaluation of Essbases schema
• May cause multiple searches.
• E.g. searching store something on
  product-store index
• Performance depends on ability to find enough
  dense dimensions.
• Efficient batch update

5
Existing methods Cont...

• Bit mapped indices
• Pros
• Low cardinality data, bit maps are both spaced
  and retrieval efficient.
• Supports bitwise operations
• Access data is clustered
• All dimensions handles symmetrically
• Cons
• Range queries
• Increased space overhead of storing the bit-maps
  specially for high cardinality data
• Expensive batch update as all bit mapped indices
  have to be modified even for a single row
  insertion

6
Existing methods... Cont

• Bit-mapped indices variants
• Compression
• Hybrid
• Dynamic Bit-maps

7
Existing methods... Cont

• Hierarchical Indices
• Example Product - Store
• Index product first also store summaries on
  product level.
• For each product value, create index for Store
  and store summaries for product-store level
• Pros
• Allows faster access to higher levels data
• Dimensions are symmetrically handled
• Cons
• Widely used index storage overhead
• The average retrieval efficiency can suffer
  because of large indexing structure

8
Existing methods Cont

• Multidimensional indices
• Use of of the indexed methods designed for
  spatial data
• E.g RTree, GridFiles etc.

9
Optimized R-Tree of OLAP data

• Rectangular dense region (only the boundaries
  that contain more than threshold number of points
• Contains a pointer to variable length array of
  (TIDs or the tuples itself)
• Points in sparse regions
• Finding dense regions
• Ask Expert?
• Use of clustering algorithm (similar algorithm
  image analysis)
• Need evaluation!!

10
R-Tree VS Bit-mapped indices

• R-Tree Pros
• Allows range queries
• Smaller space overhead
• Update is more efficient
• Bit-mapped Pros
• Faster Bit-wise operation
• Efficient for low cardinality, few restricted
  dimensions, and sparse data.

11
Conclusion

• High level overview
• Recommended readings
• MOLAP VS OLAP
• R-Tree and variants
• R-Tree alternatives
• Computational of multidimensional aggregates
• And More..