Andreas Korneliussen wrote:
Mike Matrigali wrote:
The SUR should not know anything about the underlying implementation
of the access method getting the row, so having it "read a timestamp"
from page does not work. If the timestamp is not in the rowlocation,
we could add a get a timestamp for row at this rowlocation, but forcing
two trips to the store for every row is a overhead. Rather than discuss
implementation it would be nice to understand the minimum necessary
services needed to be provided by the access method. Do the same
interfaces need to be provided by VTI's? At least
for your use I think the timestamp need only guarantee to be different
after a truncate from previous version on page.
Since you are ok with invalidating the SUR in the case of offline
compress, what about invalidating the SUR in the case of online
compress also? One way to do this is for the system catalogs to
maintain a table version number, which would be guaranteed to not
change while any sort of table intent lock was present. Any operation
which either copied rows to another container or truncated the
table would bump the version number. And holdable cursors would need
to recheck the version number after losing the lock at commit time.
I think I could go for the following solution to invalidate the SUR in
case of online compress:
- A sequence number is associated with each Container
- The sequence number is updated when doing truncate
A holdable cursor will need to reopen the controller after a commit,
since the controllers get closed at the end of the transaction (in
closeForEndTransaction(..)).
When reopening a controller, one may check that the sequence number has
not been changed since it was initially opened. If it has changed, one
can conclude that there has been a online compress, and updates cannot
be safely executed, and we may reject the reopen.
Any attempt to do update on a non-reopened controller, will fail, and a
warning given (cursor operation conflict).
This solution does not have the downside of requiring any changes to the
page layout, or RowLocation. It also does not have a cost per row. The
downside, is that a online compress will invalidate the cursor from
doing any update, even for rows which are unaffected of the truncate.
Note: the ScrollInsensitiveResultSet does not need to know anything
about the sequence number.
Andreas
This sounds like a good direction.
I was suggesting that the sequence number be maintained in the system
catalogs and owned by upper layer of the system. It seems like you are
proposing the sequence number be owned by store. If owned by store
I think I would describe the sequence number something like:
An implementation specific long which will be changed to a never
previously used number if the table undergoes a change which
results in the possibility of a RowLocation which was previously
allocated being reused in a container which was built requesting
no RowLocation reuse.
Can you explain at what point, and in which part of the code does the
system check that the sequence number has changed and then fail the
SUR? If only for SUR then there will be some querying from SUR to
store after every commit. If only in store then the closing will affect
existing holdable cursors.
The downside is that some SUR's are invalidated that didn't need to be,
but compress kicking in, in a holdable cursor in the time between a
commit and then next operation in the cursor is going to be a
rare event. The upside is that there is no extra per row overhead in
the system for the normal case.
There already exists a ddl invalidation scheme for invalidating query
plans, maybe this existing structure could be used to invalidate
SUR's after the commit?
Andreas Korneliussen wrote:
I will modify the suggestion somewhat. I think first, that offline
compress is not a problem, even for the holdable SUR. Since offline
compress moves the records to another container, the SUR cursors
should detect that container they use is no longer valid, when
renavigating to the row.
If a client of store moves a row by deleting and inserting it
somewhere else, the SUR should not find the row when trying to do
renavigate to it for update or delete, and can give an error.
What our problem is, is the case where a row is inserted into the
container, and it gets the same RowLocation as a row which we have
read into the SUR. The row which we had previously read into the SUR,
must have been deleted and purged for this to happen.
In addition, as far as I can see, for a new row to get the same
RowLocation as a row previously deleted and purged, the page for the
row, must have been truncated, and recreated.
So then how can we detect that a page has been recreated ? We could
i.e use a timestamp on the create/recreate time of the page. This
timestamp could be read by the SUR as it reads the RowLocation (so we
do not need to change the impl. of RowLocation), and again, we would
probably need to change the header for the page, so that we can store
the timestamp.
Andreas
Mike Matrigali wrote:
Some questions:
o row locations are stored in every index row. Are you proposing a
data level upgrade of every row in all databases?
o What is your proposal in the case of soft upgrade (note I believe not
supporting "holdable" SUR in soft upgrade is an option).
o The hard case is the compress case that removes pages from a file, in
this case there is no place to store the version number that you
are relying on (the same problem in the current system why truncte
can't support non-reusable rowlocations).
o Is it worth the on disk and in memory overhead to every row
location to support holdable SUR?
I believe one of the operations you are trying to address is when a
client of store moves a record by deleting and inserting it. This is
what compress does today. So if we start with row loc A pointing at
row A, and compress deletes row A and inserts it at row loc B. In both
the current and new system access to A will return an error, but
neither
will "know" that the row has been moved to a new ID. Is this ok?
If the current system always supported non-reusable row id's, even in
the truncate case do you have what you need? Again this will not
prevent clients of store from moving a row by inserting and deleting
it somewhere else.
Andreas Korneliussen wrote:
Following is a proposal to ensure that a client of store can verify
the validity of a RowLocation. A RowLocation has become invalid if
a store operation has caused it to point to another row or to a
non-existent position (deleted row or non-existing page/record-id).
I think we need a mechanism to detect that a RowLocation has become
invalid in order to implement *holdable* SUR.
To do this, I would propose:
- The RowLocation object should contain a version number for the page.
- A version number should be stored in the header for a Page
- Whenever an operation which may invalidate row-locations is
executed, the version number for the page is updated. These
operations include online/offline compress.
- When navigating to a RowLocation which has invalid version
number, the store may fail (i.e return false)
The page header for a stored page, currently has a number of fields
which are intended for future use, and it seems that it is possible
to use these fields without breaking backward compatibility.
I noticed one of the fields in the header is named "generation"
(from StoredPage.java):
* 4 bytes integer generation generation number of
this page(FUTURE USE)
* 4 bytes integer prevGeneration previous generation of
page (FUTURE USE)
Could I use the generation field for this, or has it been reserved
for something else ? Alternatively, I could use one of the other
long fields reserved for future use.
Any comments ?
Thanks
--Andreas
