directory-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Selcuk AYA <>
Subject Re: Txn Work Update(2)
Date Thu, 15 Dec 2011 11:25:05 GMT
Hi all,

update on my status:

* There is going to be a big change in my life soon and I will do a
big move. So I will be MIA for some time( around 2 months) and after
that I will have less time to contribute but I will try to keep being
involved in development and discussions.

* I did a recent check in. There are some tests that fail. I will list
them below. The basic reason that those tests are failing is that they
do queries and updates of which  some are transactional, some are not
and this leads to wrong results. And the basic reason they do queries
and updates in a non transactional way is that I did not put
transaction begin/commif/abort( ie, txn demarcation) at CoreSession.
This is the layer we start the execution of  the basic ldap operations
but I thought there might be clients of CoreSession which want to
group operations together and execute them as a single txn(If you
think this is not necessary, please let me know). Also, there is at
least search, for which the cursor is kept open above CoreSession, So
for search, the caller has to handle txn commit, abort.Also there
might be some cases where operations do not use session at all.

As a result, I decided to handle each  caller of CoreSession
operations separately and this increased the number changes that had
to be done. I handled ldap protocol handlers and change password.
Right now StockServerSuite is passing except three tests.
StockCoreSuite and MigratedStockCoreSuite is failing. The reason is
they use LDAPCoreConnection and this is not transactional yet.

For anybody who wants to get involved, here is the list of todo:
* Figure out why SaslBind and DelegatedAuthIT in stockserver suite is
failing. We get a KRBException here and I do not know this module well

*Handle LdapCoreConnection. This is used in applications which embed
apacheds i guess? The basic difficulty is that users of this class get
a cursor and the transaction should be comitted or aborted only when
this cursor closes . Moreover, when a cursor is open, user of
ldapcoreconnection can call other ldap operations and it might make
sense to group these in a transaction. If this is true, then it should
be the responsibility of the callers of LdapCoreSession to start and
end txns. If we are not going to let the user do this, we need to
store a pointer to txn context in the cursor and end the txn when
cursor is closed taking into account that the txn might have to be
upgraded to a readwrite txn from a read only txn if the user does
changess while holding the cursor open.

* There are some tests which call session operations directly, these
should be wrapped in txn begin/end.

*I put some tests to AVL and jdbm partitions to test txns without the
effect of interceptors and logical caches. These are multithreaded and
are good to test the robustness of the data change and query layer and
txn subsystem. Add more tests like this.

* Handle logical caches. As we talked handle these with read write
locks taking cleanup in  case of an aborted txn into account.

Please let me know if anyone wants to attack any of these. I might
provide more input if needed.


On Thu, Dec 1, 2011 at 10:47 AM, Emmanuel Lecharny <> wrote:
> On 12/1/11 8:37 AM, Selcuk AYA wrote:
>> On Wed, Nov 30, 2011 at 3:26 PM, Emmanuel Lecharny<>
>>  wrote:
>>> On 11/30/11 11:48 AM, Selcuk AYA wrote:
>>>> Hi all,
>>> hi !
>>>> time to give an update on the txn status. I just checked in some code
>>>> with explanation of the changes for each project separately. Here is
>>>> just an overall summary and todo:
>>>> *I added txn being/commit/abort to ldap requests and ran some test
>>>> with txns enabled. Put in various fixes to the txn layer based on the
>>>> tests. So first signs life signs for txns. More tests will be done but
>>>> I need to digress into AVL first. I also need to handle paged search.
>>> Great !
>>> Paged search is just a a normal search, but one that last longer. The way
>>> we
>>> handle it is by using a counter of sent responses.
>>> Basically, entries are sent one by one to the client, using a
>>> SearchResultEntry to hold the entry. With the pagedSearch, once we have
>>> reached the limit set by the client, we stop the sending of entries, and
>>> hold the cursor in the user's session, so that he can send back a new
>>> search
>>> request for then next entries - if needed -. Is there anything special
>>> you
>>> need to deal with here ? (btw, the cursor will be deleted as soon as the
>>> session will be closed. I hope this cleanup is effective...)
>> There is not much todo here as you guessed. The only problem is the
>> transaction context is stored in a thread local variable. When paged
>> search send partial data, we should store the context somewhere and
>> hand it off to the thread that will continue processing the search.
> ThradLocal is ok during the duration of the request. Why don't we use a
> session context that is associated with the user ? Right now, we are using a
> Context for each request, but inside it, we store the IoSession, where we
> can store some user's specific information, instead of storing them in a
> ThreadLocal.
>>>> *I have also been working on removing interfaces related to
>>>> modification from the partition
>>> Hmmm, can you be a bit more specific ?
>> When I say modification or change, I mean
>> modify/move/rename/moveAndRename/add/delete. Instead of partitions
>> implementing these separately, there is a layer above partitions which
>> handle these operations transactionally. The class that handles this
>> is called OperationExecutionManager. Partitions expose master table
>> and indices for the operation execution manager execute the
>> operations.
> Ok, understood. Fine with me. It will help a lot when it comes to write a
> new partition, instead of having to implement all the logic for each of
> these operations. Good move.
>>>> and have been trying to get rid of the
>>>> Store interface. Got rid of the store interface except for config
>>>> methods and calls from the ui layer. I removed the modifications up
>>>> but still have to call the modification methods of the partition
>>>> interface because SchemaPartition triggers some changes based on the
>>>> logical  change. I believe those changes can be removed to schema
>>>> interceptor?
>>> We can't modify a schema. We can only add or delete a schema entry. This
>>> is
>>> a limitation, and I think it's mandated by the spec (don't remember where
>>> it
>>> comes from, but the idea is that modifying a schema element might imply a
>>> global modification on all the entries, something we don't want to
>>> accept).
>>>> In general, if a partition wants to do a change based on
>>>> a data update, then it should be able to do it during master table
>>>> update or by registering an interceptor(For example I changed multi
>>>> file ldif partition to do file updates when master table is updated).
>>>> Then we can get rid of the modification methods on partitions.
>>> you mean, the modify/move/rename/moveAndRename methods ?
>> As mentioned above, by change or modification, I mean any operation
>> that ends up in update of data. For schema partitions, I believe
>> changes triggered after logical ldap change operations can be moved to
>> an interceptor.
> For schema, we now update the LDIF partition, whch means we go down to the
> partition layer, and everything else is managed by the interceptor, already
> (at least in trunk, and I think I have merged it back to your branch. Will
> check today).
>>>> * Would be good to have statement snapshot isolation for
>>>> EntryFilteringCursor. Not having this doesnt affect ldap requests but
>>>> might affect code that works with EntryFilteringCursor directly.
>>>> Currently there are few cases of this especially during startup.
>>> I'll buy some more explanation : what is a 'statement snapshot' ?
>> Say you do a search and get an entry filtering cursor and then you
>> modify the entries that you get from the cursor. Something like this:
>> beginTransaction()
>> cursor = ..
>> while ( )
>> {
>>   entry = cursor.get();
>>   modify entry.
>> }
>> endTransaction()
>> If we have statement snapshot for the search, cursor will not see any
>> updates to the master table of index tables done through the
>> modifications.  This might be useful in some cases.
> Ok, got it.
> --
> Regards,
> Cordialement,
> Emmanuel Lécharny

View raw message