[
https://issues.apache.org/jira/browse/PHOENIX-4550?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16338493#comment-16338493
]
James Taylor commented on PHOENIX-4550:
---------------------------------------
I think there's a theoretical problem with updatable views in general. There
could be multiple views for the same row. This is arguably a situation we may
want to prevent, but we're not doing that today. For example, say you have the
following hierarchy:
T (A, B, C)
V1 (D, E) FROM T WHERE A = 1
V2 (F, G) FROM T WHERE A = 1 and B = 2
The same rows in table T could be in both V1 and V2. So then T would occupy
positions 1-3, V1 would occupy positions 4-5, and V2 would occupy positions
6-7. Depending on which view you updated through, you'd have nulls in either
positions 4-5 or 6-7. In cases like this, there's no advantage to having a
mapping or declaring the max number of columns.
If we detect this and disallow it at creation time, we can pursue this JIRA.
I'll file a separate JIRA for that. In reality, we don't have use cases in
which view rows overlap, so this is kind of a theoretical problem.
So assuming the views aren't overlapping, how would you deal with columns that
have been dropped? Also, are you thinking to push this map through every
SingleCellColumnExpression? Wouldn't that be expensive, especially if there are
many columns and many column references in a query?
With the alternative, preallocating a fixed number of columns, you'd need to
push the preallocated number plus the original starting column qualifier of a
view to figure out the array position. The downside is that the preallocated
columns would be wasteful.
Not sure that the map idea solves the issue of when a column is added to a base
table since it needs to be in the same array position for all rows.
> Allow declaration of max columns on base physical table
> -------------------------------------------------------
>
> Key: PHOENIX-4550
> URL: https://issues.apache.org/jira/browse/PHOENIX-4550
> Project: Phoenix
> Issue Type: Improvement
> Reporter: James Taylor
> Priority: Major
>
> By declaring the max number of columns on a base table, we can optimize the
> storage for SINGLE_CELL_ARRAY_WITH_OFFSETS by not storing null values for the
> columns preceding the initial column of a view. This will make a huge
> difference in storage when you have a base table with many views. For example:
> {code}
> -- Declare that the base table will have no more than 10 columns
> CREATE IMMUTABLE TABLE base (k1 VARCHAR, prefix CHAR(3) v1 DATE,
> CONSTRAINT pk PRIMARY KEY (k1, prefix))
> MULTI_TENANT = true,
> MAX_COLUMNS = 10;
> CREATE VIEW v1(k2 VARCHAR PRIMARY KEY, v2 VARCHAR, v3 VARCHAR)
> AS SELECT * FROM base WHERE prefix = 'A00';
> CREATE VIEW v2(k2 VARCHAR PRIMARY KEY, v2 VARCHAR, v3 VARCHAR);
> AS SELECT * FROM base WHERE prefix = 'A10';
> ...
> {code}
> As the number of views grow, the difference between the base table column
> encoding (column #1) and the starting column number of the view (since the
> starting offset is determined by an incrementing value on the base table)
> will increase. This bloats the storage as we need to store null values for
> column encodings between the base table column and the starting column of the
> view.
> Instead, we'll pass through the MAX_COLUMNS value for queries and anything
> column encoding less than this we know it'll be at the start. Anything
> greater and we'll start the search from <column encoding> - <minimum view
> column encoding>.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
