Uploaded image for project: 'MariaDB ColumnStore'
  1. MariaDB ColumnStore
  2. MCOL-3574

EMPTY column for persistant storage

    XMLWordPrintable

Details

    Description

      MCS 6.3.1 has a number(column type dependent) special column values(magic values) for empty records. This allows in-place DELETE-s but slows down DELETE-s a lot b/c MCS needs to write magics into all columns for a record deleted.
      This project describes another approach with an additional auxiliary column(every table will have such column) that stores EMPTY attribute as its Nth bit. With such record attribute MCS can:

      • improve INSERT cache speed. MCS inserts/marks the record deleted until commit arrives.
      • improve DELETE speed/user experience. MCS journals/changes a single-bit per-record and not N * columns number per record.
      • improve UPDATE speed. MCS becomes append-only so that the whole UPDATE will become an INSERT + DELETE operation.

      These are immediate benefits. There are some potential positive side-effects, e.g. MCS should be able to cope with medium OLTP workload that it can't do previously so that OLTP to OLAP replication can be done.

      The first cut of the implementation plan consists of a number of steps:

      • existing installations migration planning(this will affect the feature implementation)
      • hidden aux column creation running CREATE TABLE in DDL or plugin. The aux column data type empty value signifies empty record so that EMPTY value code can be reused for this column.
      • basic scanning support in EM or PP
      • basic INSERT support(a record is on or off by default) in WE
      • basic bulk insert support in WE
      • basic DELETE support via an aux column) in WE

      Here are some notes on the bullets above.
      Migration can be seamless or with additional upgrade downtime. The first approach reduces upgrade downtime but the performance will be potentially degraded b/c MCS now needs to detect whether aux column exists or not and use specific code for reads/inserts/deletes/updates. Developer's needs also to maintain both processing paths that adds complexity. The second migration approach is to run a binary or update routine in the cluster that will add/populate aux column for every table. We should provide customers with ways to rollback update changes on per-table basis.
      I presume that at any moment MCS knows table OID so we can make aux column OID so column OID is table OID + 1.
      Basic scanning support can be implemented adding aux column filter for any disk-based scan so there is no need to change actual scanning code.
      There are two options changing INSERT path. The most simple is to avoid any changes so that MCS WE paths will work with the table + additional column with its EMPTY magic. The more elaborate is to reduce the number of blocks we put into a VersionBuffer cache to one, namely aux column only. If ROLLBACK fires MCS needs to restore only aux column block but not all other columns' blocks. The best thing about this particular part is that VBBM and VSS will be used for aux column only but not for all table columns.
      Bulk insert operation must extend aux column and populate it with EMPTY values.
      DELETE operation will flip empty record attribute bit value.

      Attachments

        Issue Links

          Activity

            People

              tntnatbry Gagan Goel (Inactive)
              LinuxJedi Andrew Hutchings (Inactive)
              Votes:
              0 Vote for this issue
              Watchers:
              2 Start watching this issue

              Dates

                Created:
                Updated:
                Resolved:

                Git Integration

                  Error rendering 'com.xiplink.jira.git.jira_git_plugin:git-issue-webpanel'. Please contact your Jira administrators.