On Wed, May 31, 2017 at 3:02 PM, Shubham Barai <shubhambara...@gmail.com> wrote:

> I have been accepted as GSoC student for the project "Explicitly support
> predicate locks in index access methods besides b-tree". I want to share my
> approach for implementation of page level predicate locking in gist index.

For the benefit of all following the discussion, implementing
support in an index AM conceptually consists of two things:
(1) Any scan with user-visible results must create SIRead predicate
locks on "gaps" scanned.  (For example, a scan just to find an
insertion spot for an index entry does not generally count, whereas
a scan to satisfy a user "EXISTS" test does.)
(2) Any insert into the index must CheckForSerializableConflictIn()
at enough points that at least one of them will detect an overlap
with a predicate lock from a preceding scan if the inserted index
entry might have changed the results of that preceding scan.

Detecting such a conflict does not automatically result in a
serialization failure, but is only part of tracking the information
necessary to make such a determination.  All that is handled by the
existing machinery if the AM does the above.

With a btree index, locking gaps at the leaf level is normally
sufficient, because both scan and insertion of an index entry must
descend that far.

> The main difference between b-tree and gist index while searching for a
> target tuple is that in gist index, we can determine if there is a match or
> not at any level of the index.

Agreed.  A gist scan can fail at any level, but for that scan to
have produced a different result due to insertion of an index entry,
*some* page that the scan looked at must be modified.

> The simplest way to do that will be by inserting a call for
> prdicatelockpage() in gistscanpage().

Yup.

> Insertion algorithm also needs to check for conflicting predicate locks at
> each level of the index.

Yup.

> We can insert a call for CheckForSerializableConflictIn() at two places in
> gistdoinsert().
>
> 1. after acquiring an exclusive lock on internal page (in case we are trying
> to replace an old search key)
>
> 2. after acquiring an exclusive lock on leaf page
>
> If there is not enough space for insertion, we have to copy predicate lock
> from an old page to all new pages generated after a successful split
> operation. For that, we can insert a call for PredicateLockPageSplit() in
> gistplacetopage().

That all sounds good.  When you code a patch, don't forget to add
tests to src/test/isolation.

Do you need any help getting a development/build/test environment
set up?

--
Kevin Grittner
VMware vCenter Server
https://www.vmware.com/


-- 
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers

Reply via email to