Uploaded image for project: 'Derby'
  1. Derby
  2. DERBY-6921

How good is the Derby Query Optimizer, really



    • Improvement
    • Status: Open
    • Minor
    • Resolution: Unresolved
    • None
    • None
    • SQL


      At the 2015 VLDB conference, a team led by Dr. Viktor Leis at Munich
      Technical University introduced a new benchmark suite for evaluating
      database query optimizers: http://www.vldb.org/pvldb/vol9/p204-leis.pdf

      The benchmark test suite is publically available:

      The data set for running the benchmark is publically available:

      As part of Google Summer of Code 2017, I am volunteering to mentor
      a Summer of Code intern who is interested in using these tools to
      improve the Derby query optimizer.

      My suggestion for the overall process is this:
      1) Acquire the benchmark tools, and the data set
      2) Run the benchmark.
      2a) Some of the benchmark queries may reveal bugs in Derby.
      For each such bug, we need to isolate the bug and fix it.
      3) Once we are able to run the entire benchmark, we need to
      analyze the results.
      3a) Some of the benchmark queries may reveal opportunities
      for Derby to improve the query plans that it chooses for
      various classes of queries (this is explained in detail in the
      VLDB paper and other information available at Dr. Leis's site)
      For each such improvement, we need to isolate the issue,
      report it as a separable improvement, and fix it (if we can)

      While the benchmark is an interesting exercise in and of itself,
      the overall goal of the project is to find-and-fix problems in the
      Derby query optimizer, specifically in the 3 areas which are
      the focus of the benchmark tool:
      1) How good is the Derby cardinality estimator and when does
      it lead to slow queries?
      2) How good it the Derby cost model, and how well is it guiding
      the overall query optimization process?
      3) How large is the Derby enumerated plan space, and is it

      While other Derby issues have been filed against these questions
      in the past, the intent of this specific project is to use the concrete
      tools provided by the VLDB paper to make this effort rigorous and
      successful at making concrete improvements to the Derby query

      If you are interested in pursuing this project, please take these
      considerations into mind:
      1) This is NOT an introductory project. You must be quite familiar
      with DBMS systems, and with SQL, and in particular with
      cost-based query optimization. If terms such as "cardinality
      estimation", "correlated query predicates", or "bushy trees"
      aren't comfortable terms for you ,this probably isn't the
      project you're interested in.
      2) If you are new to Derby, that is fine, but please take advantage
      of the extensive body of introductory material on Derby to
      become familiar with it: read the Derby Getting Started manual,
      download the software and follow the tutorials, read the documentation,
      download the source code and learn how to build and run the
      test suites, etc.
      3) All I have presented here is an *outline* of the project. You will
      need to read the paper(s), study the benchmark queries, and
      propose a detailed plan for how to use this benchmark as a tool
      for improving the Derby query optimizer.

      If these sorts of tasks sound like exciting things to do, then please
      let us know!




            harshvardhan145 Harshvardhan Gupta
            bryanpendleton Bryan Pendleton
            0 Vote for this issue
            3 Start watching this issue



              Time Tracking

                Original Estimate - 2,016h
                Remaining Estimate - 2,016h
                Time Spent - Not Specified
                Not Specified