STREAM: The Stanford Data Stream Management System

1
STREAM The Stanford Data Stream Management System

• Rebuttal Team
• Mingzhu Wei
• Di Yang
• CS525s - Fall 2006

2
Rebuttal Areas

• Foundation
• Windows
• Joins
• Full Recalculation Strategy
• Language Issues

3
Foundation

• Rebuttal
• No proof provided to guarantee correctness or
  completeness of query plans resulting from the
  combination of operators, queues, and synopses
• Analysis
• STREAM is based on relation database theory
• Relational databases have been around for a long
  time
• Proofs exist that demonstrate their correctness
• CQL is a minor extension to SQL
• More effort could have been put into providing a
  more formal proof of CQL

4
Windows

• Rebuttal
• Stream does not provide value-based windows
• For example, without a value-based window the
  system cannot process a query such as
• Give me the name of the students who have the top
  10 exam scores, efficiently
• Analysis
• Feature not supported by STREAM

5
Joins

• Rebuttal
• STREAM only uses the self-purge mechanism when
  performing a window-based join
• Analysis
• STREAMs criteria for judging when a tuple (in
  the state of the window) has expired is
  determined by comparison of its timestamp with
  that of the new incoming tuples in the same
  stream
• Cross-purge might be more efficient in some cases
• Cross-purge Compare timestamps across two
  streams

6
Full Recalculation Strategy

• Rebuttal
• Stream uses a full recalculation strategy for
  result updating
• Could be very inefficient with big window sizes
• Example
• We are trying to join two windows each of size
  1000
• If both windows only slide 10 at each time,
  recalculation for the whole result would be much
  more expensive, than incremental result updating
• Analysis
• Using an Incremental Result Update Strategy might
  be more efficient in some cases
• Keep most of the joined result and only calculate
  those for newly arrived tuples

7
Language Issues

• Rebuttal
• Stream does not provide the stream to stream
  operator
• Analysis
• The absence of a stream-to-stream operator is not
  explicitly justified in the paper
• Its absence is reasonable because STREAM
  operators treat all input as relations
• STREAM does provide operators for converting
  streams to relations and for converting relations
  to streams

8
Language Issues (cont)

• Rebuttal
• Stream uses an append-only model
• It does not provide an operator for updating data
  value in stream
• Analysis
• Although not perfect, this is a common assumption
  in current stream processing papers

9
Conclusion

• Foundation
• Windows
• Joins
• Full Recalculation Strategy
• Language Issues