Uploaded image for project: 'Derby'
  1. Derby
  2. DERBY-1506

full table scans of tables which don't use indexes, which have blobs, but don't reference blob data still read all pages of the table



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


      A full table scan that does not use an index always reads every allocated page of the table through the cache. In two cases
      this means logically it reads more data that necessary: long rows (where the head columns span multiple pages) and long columns
      (columns where the data in a single column spans multiple pages). In both these cases the space for the "overflow" portion of the
      row or column is currently stored in the same space as the regular "main" pages. The current implementation of a table scan of
      a heap container is to call raw store to give it a linear list of main pages with rows, raw store conglomerate implementations step through each allocated page in the container and returns the "main" pages (reading the overflow pages into cache, identifying them, and skipping them)
      the access layer which then returns rows as requested to the query processing layer.

      If a table contains rows with very long columns (ie. 2gig blobs), and the tablescan does not request the blob data then a lot of data
      is read from disk but not required by the query.
      A more unusual case is a table scan on requiring a few columns from a table made up of 2 gig rows made up of all less than 32k columns.,
      in this case also derby will read all pages as part of a tablescan even if only the first column is the only required column of the chain.

      Note that this is only a problem in tablescan of heap tables. In both cases if an index is used to get the row, then ONLY the required data is
      read from disk. In the long column case the main row has only a pointer to the overflow chain for the blob and it will not be read unless the
      blob data is required. In the long row case data, columns appear in the container in the order they are created in the original "create table"
      statement. Data is read from disk into cache for all columns from the 1st up to the "last" one referenced in the query. Data objects are only
      instantiated from the cache data for the columns referenced in the query.

      I have marked this low in priority as I believe that full, unindexed tables scans of tables with gigabyte blobs are not the normal case. Seems like most applications would do keyed lookups of the table. But there may be apps that need to
      do full table reporting on the non'-blob data in such a table.


        Issue Links



              Unassigned Unassigned
              mikem Mike Matrigali
              4 Vote for this issue
              5 Start watching this issue