Uploaded image for project: 'Cassandra'
  1. Cassandra
  2. CASSANDRA-9415

Implicit use of Materialized Views on SELECT

    XMLWordPrintableJSON

Details

    • Improvement
    • Status: Resolved
    • Normal
    • Resolution: Later
    • None
    • None

    Description

      CASSANDRA-6477 introduces Materialized Views. This greatly simplifies the write path for the best-practice of "query tables". But it does not simplify the read path as much as our users want/need.

      We suggest to folks to create multiple copies of their base table optimized for certain queries - hence "query table". For example, we may have a USER table with two type of queries: lookup by userid and lookup by email address. We would recommend creating 2 tables USER_BY_USERID and USER_BY_EMAIL. Both would have the exact same schema, with the same PRIMARY KEY columns, but different PARTITION KEY - the first would be USERID and the second would be EMAIL.

      One complicating thing with this approach is that the application now needs to know that when it INSERT/UPDATE/DELETEs from the base table it needs to INSERT/UPDATE/DELETE from all of the query tables as well. CASSANDRA-6477 covers this nicely.

      However, the other side of the coin is that the application needs to know which query table to leverage based on the selection criteria. Using the example above, if the query has a predicate such as "WHERE userid = 'bhess'", then USERS_BY_USERID is the better table to use. Similarly, when the predicate is "WHERE email = 'bhess@company.whatever'", USERS_BY_EMAIL is appropriate.

      On INSERT/UPDATE/DELETE, Materialized Views essentially give a single "name" to the collection of tables. You do operations just on the base table. It is very attractive for the SELECT side as well. It would be very good to allow an application to simply do "SELECT * FROM users WHERE userid = 'bhess'" and have that query implicitly leverage the USERS_BY_USERID materialized view.

      For additional use cases, especially analytics use cases like in Spark, this allows the Spark code to simply push down the query without having to know about all of the MVs that have been set up. The system will route the query appropriately. And if additional MVs are necessary to make a query run better/faster, then those MVs can be set up and Spark will implicitly leverage them.

      Attachments

        Issue Links

          Activity

            People

              Unassigned Unassigned
              brianmhess Brian Hess
              Votes:
              5 Vote for this issue
              Watchers:
              11 Start watching this issue

              Dates

                Created:
                Updated:
                Resolved: