[
https://issues.apache.org/jira/browse/DERBY-2212?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12535561
]
Anurag Shekhar commented on DERBY-2212:
---------------------------------------
1) Current code changes B2I so needs upgrade code or else no old database with
any index will be able to be read.
I do have code for soft and hard upgrade in my todo list. I will create an jira
for the same.
2) The approach seems complicated and error prone to me. Treating compare of
nulls one way through an extra overloaded arg sometimes and sometimes seems
overly complicated.
It is just one of the possible approach I am trying it out only because I felt
its a good solution. I am open to take other approach.
3) I need to think about it more carefully to understand how treating compare
of nulls differently in the same tree might affect things like logical undo
during restart recovery after an uncommitted insert where the row may have
moved drastically since the insert do to lots of other inserts.
I will write some tests related to this scenario and get back.
4) before commit needs a bunch of null heavy additional test cases.
Yes I need to write several test cases for new behaviour and also some make
sure its not breaking something else in some other part of the code base.
1) optimizer probably will assume 1 row from unique index in this case, now it
may have to use some other estimate when dealing with nulls. If the index was
non-unique rather than unique current code would just work.
Optimizer has a provision to treat IS NULL differently from '='. I am using
that in my current patch and it appears to be working fine for very basic tests.
2) unique index locking isolation level algorithm is different from duplicate
index locking isolation level for serializable. It may have to now change.
Previously a previous key lock is not necessary to protect a "phantom" insert
range for a unique index - after your change maybe it is? Again if the index
was marked non-unique at implementation level then current code would just
work.
I need to check about it. I will test and debug the code in multi user mode and
get back.
> Add "Unique where not null" to create index
> -------------------------------------------
>
> Key: DERBY-2212
> URL: https://issues.apache.org/jira/browse/DERBY-2212
> Project: Derby
> Issue Type: Improvement
> Components: SQL
> Affects Versions: 10.2.1.6
> Reporter: Oleksandr Alesinskyy
> Assignee: Anurag Shekhar
> Attachments: derby-2212preview.diff, derby-2212preview2.diff
>
>
> Derby prohibits creation of unique constraints on nullable colums (as well if
> only some columns in the constraint list are nullable) and treat nulls in
> unique indexes as normal values (i.e. only one row with null values in
> indexed columns may be inserted into the table). This bahavior is very
> restrictive, does not completely comply with SQL standards (both letter and
> intent) as well as with business needs and intending meaning of NULL values
> (2 null values are not considered as equal, this comparision shall return
> NULL, and for selection criteria boolean null is treated as FALSE).
> This behavior, as far as I can see, is modelled after DB2 (and differs from
> behavior of most other major databases, like SyBase, Oracle, etc.).
> But even DB2 provide some means to alleviate these restrictions, namely
> "UNIQUE WHERE NOT NULL" clause for CREATE INDEX statement.
> It will be very good if such "UNIQUE WHERE NOT NULL" clause will be
> introduced in Derby.
> Regards,
> Oleksandr Alesinskyy
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
