[ApacheDS] Implementing isolation using multi-version concurrency control (MVCC)

[Alex Karasulu <ak...@apache.org>]

Hi all,

Emmanuel and I were having an interesting conversation about the kinds of
transaction properties needed by ApacheDS to comply with various
specification requirements. For example all LDAP operations must be atomic
and must preserve consistency. Furthermore, one can easily debate the need
for isolation where any operation does not see or get impacted by the
partial changes of another concurrent operation.

We started discussing these ACID properties and ways in which they could be
implemented. Isolation is very interesting now thanks to directory
versioning and the change log. We can easily implement isolation now. When a
relatively lengthy operation like a search is being conducted, it should not
see modifications within scope that occur after the search began. The search
operation in the example can be replaced with any other operation minus all
unbind, some extended, and all abandon requests.

The change log, not only tracks each change, but it allows the directory
server to inject the "revisions" attribute into entries. The revisions
attribute is multi-valued and lists all the revisions of changes which
altered the entry. For the search example, we can conduct the search
operation while locking it down to a revision.  This is best implemented by
conditionally filtering out or injecting candidates with revisions greater
than the revision at which the search operation started. Let's call the
revision when search started S. So entries in the server which posses
revision numbers greater than S need further evaluation. We have to evaluate
if the filter matches these entries with revisions > S when their state was
at revision S. This may require some reverse patching and re-evaluation of
the filter on the patched entry in state S.  This is not so bad because
there usually are not that many changes taking place at the same time on the
same entry: meaning the number of LDIF's to patch on an entry to evaluate in
it's former state at S will be small. This way we effectively lock down the
search to a specific revision, giving the search operation what appears to
be a snapshot of the DIT. The search results will not be impacted by any
concurrent changes.

Again similar strategies can be applied to any other operation in place of
the search example with specific consideration for each operation.

This is a form of multi-version concurrency control [0], rather than using
locking mechanisms to achieve isolation. This technique works pretty well in
modern DBMS'.

--Alex

--------
[0] - http://en.wikipedia.org/wiki/Multiversion_concurrency_control

Re: [ApacheDS] Implementing isolation using multi-version concurrency control (MVCC)

[Alex Karasulu <ak...@gmail.com>]

On Sat, Jan 31, 2009 at 5:25 AM, Emmanuel Lecharny <el...@apache.org>wrote:

> Alex Karasulu wrote:
>
>> Hi all,
>>
>> Emmanuel and I were having an interesting conversation about the kinds of
>> transaction properties needed by ApacheDS to comply with various
>> specification requirements. For example all LDAP operations must be atomic
>> and must preserve consistency. Furthermore, one can easily debate the need
>> for isolation where any operation does not see or get impacted by the
>> partial changes of another concurrent operation.
>>
>> We started discussing these ACID properties and ways in which they could
>> be
>> implemented. Isolation is very interesting now thanks to directory
>> versioning and the change log. We can easily implement isolation now. When
>> a
>> relatively lengthy operation like a search is being conducted, it should
>> not
>> see modifications within scope that occur after the search began. The
>> search
>> operation in the example can be replaced with any other operation minus
>> all
>> unbind, some extended, and all abandon requests.
>>
>>
> Atomicity and Isolation are both complex to guarantee in a LDAP server.
>
> If we think about Atomicity, for instance, even if we can guarantee it
> somehow for somesimple operation like Modify, Add or Delete, for the ModDN
> operation is not that simple. We have to guarantee that all the potential
> renames are done - or reverted - as a whole.


Yep. I was thinking about this exact example and how the modifyDn operation
could potentially change a lot of descendants.


> As this operation can impact a big part of the server, and could take
> several seconds (minutes, hours, dependening on the number of entries), this
> is obviously not trivial. However, moddn operation aren't the most frequent
> one. Regarding the most simpler operation (add, delete and modify), I think
> we should implement some kind of "transaction" in the backend : the modified
> entry has to be tagged as 'under modification' until the backend has updated
> correctly the modification (or rollbacked it).


Well I'd like to avoid using the partition to implement something like this
since it is not a generalized solution.  When this capability is pushed into
a partition then all implementations will need to implement it their own
way. Partitions are already complicated to implement so adding this
requirement on top can make it overwhelming.  We want people to be able to
make partitions for whatever data they want to present through LDAP to
create a rich environment for choosing different partitions.

For example JDBM offers transactions. We could have piggy backed on this
feature to provide ACID properties under JDBM. Their are some quirks though
with how JDBM does this but it can be leveraged.  I decided a while back
while implementing the partition not to do this because I wanted the
transaction management to be handled above the partitions.


> Then we can untag the entry, and it's available back. How the CL can come
> into play here is to be discussed. IMHO, the CL and this 'transaction' will
> work hand to hand at some point.
>

Yes although I don't know how yet I agree that the CL is key here.  Perhaps
we need to build a transaction manager on top of the CL that handles and
tracks these things as well as manages the transactions. The tracking of a
transaction context, it's commit, or rollback is all up to a transaction
manager.  The CL is just there for history and logging and is something the
transaction manager uses as a tool to do it's job.


>
> Regarding isolation, it's a bit more difficult, as a search can already
> have sent back some results which could be change by another modify
> operation. This is especially the case for a ModDN operation.
>
>> The change log, not only tracks each change, but it allows the directory
>> server to inject the "revisions" attribute into entries. The revisions
>> attribute is multi-valued and lists all the revisions of changes which
>> altered the entry. For the search example, we can conduct the search
>> operation while locking it down to a revision.
>>
> That does not work for deleted entries, obviously ...
>

Sure it can work for deletes.  You must have wrote this without reading
further.  Basically according to the change log, any entries with changes
after time S when search began are evaluated to see if they should be
included or removed from the result set returned by the search operation.
You don't just let search produce candidates returning all and forgetting
about the changes that occurred after the search began.


>
>   This is best implemented by
>> conditionally filtering out or injecting candidates with revisions greater
>> than the revision at which the search operation started. Let's call the
>> revision when search started S. So entries in the server which posses
>> revision numbers greater than S need further evaluation. We have to
>> evaluate
>> if the filter matches these entries with revisions > S when their state
>> was
>> at revision S. This may require some reverse patching and re-evaluation of
>> the filter on the patched entry in state S.  This is not so bad because
>> there usually are not that many changes taking place at the same time on
>> the
>> same entry: meaning the number of LDIF's to patch on an entry to evaluate
>> in
>> it's former state at S will be small. This way we effectively lock down
>> the
>> search to a specific revision, giving the search operation what appears to
>> be a snapshot of the DIT. The search results will not be impacted by any
>> concurrent changes.
>>
>>
> Well, I don't think this is the best approach. In any case, a Ldap Search
> is considered as a dirty read. We have no way to lock down the modification
> on the read entries. So the user has _no_ guarantee that the entry he gets
> back will be valid. Usually, it doesn't matter, because the ratio of read vs
> writes on a LDAP server is just so big that we consider we don't have
> modifications. So we can simply return the latest revision, whatever it is.
> Anyway, there is another aspect we have to consider : once the user gets his
> entry, and before he uses it, before potentially send it back as a
> modification to the server, the very same entry can have been modified in
> between. As we don't lock entries, we can't protect the users from such a
> case.


The point is, the search should be conducted as if it were performed on a
snapshot of the DIT at the time the search was issued.  Now if you get dirty
reads then you can have inconsistent views of the information which is rare
yes but possible.

The beauty of this approach is that it allows us to search the DIT at an
time not just when the search occurred (time S as stated above).  This
actually allows us to for example to implement a "search in the past"
control where the user can conduct a search on the server at an earlier
time.  This is insanely powerful in terms of being able to have versioned
views in the server of information.  Say you have some application's
configuration in the DIT and you locked the appilcation down to read at some
time/rev associated with a release date for the application.  So the
application pulls it's configs from the DS with this control. Then you
continue making changes to the configuration data as you change the
application in developement. You can setup and run both the production
application and the development application but using different revisions
when performing the lookups.

The power this provides for configuration management is awesome to say the
least.  Now I am digressing.


>
>
> <digression>
> We have to remember here that we are not dealing with bank accounts and
> balances, or nuclear plants. Most of the case, we are using a LDAP server to
> handle identities. They rarely change, or when they do, it's because the
> person owning this identity is changing his own identity - thus limitating
> the odds that he is using it at the same time -. Usually, if we think about
> authorizations, which are subject to way more changes that identities, we
> can't consider it as a continuous flow of modification.


Yes yes I think you know that I would know this :-). However we're just
thinking about these features.  We have to know how much of a bitch it is by
going into the details then stepping back and looking at the big picture.
Finished reading the digression and I think we've told each other this
several times before in the past.

I used to think this way a lot but there are several things that have
impacted my views.  For example we need atomicity period which is a bigger
deal with moddn.  Also with triggers we need atomicity badly too.  These are
protocol requirements and there's no arguing with that. To do all this we're
going to need a transaction manager that can begin, commit or rollback sets
of changes to the DIT.


>
>
> In other words, creations or deletion of entries might be frequent,
> modification should be quite rare.


In today's day and age directories are being used in a myriad of ways -
especially after the identity buzz. I don't think we can be as comfortable
generalizing the scenario to expect as much as we used to.


>
>
> I have gathered some stats from some of my clients, and the ration
> changes/reads is like 1/5000... I would be interested to get more numbers
> here !
>
> Last, not least, I consider that if the ratio goes up to something like
> 1/100, then it's time to consider using a transactional system, namely, a
> RDBMS.


Yeah stats are good but we need a lot of them from different industries to
understand this picture.


>
> </digression>
>
> So far, I'm not saying that it's wrong to think about using a MVCC system,
> but I'm a bit sceptic about the gain in term of isolation (the I in ACID) it
> will offer in our case.
>

We're doing it somewhat today just not going all the way. This whole
versioning facility is the basis for it.  I don't think I understand why
you're skeptical. I hope to explore how this sense is coming to you because
maybe you have a seriously important reason behind it and just are not
vocalizing it.

Thanks,
Alex

Re: [ApacheDS] Implementing isolation using multi-version concurrency control (MVCC)

[Emmanuel Lecharny <el...@apache.org>]

Alex Karasulu wrote:
> Hi all,
>
> Emmanuel and I were having an interesting conversation about the kinds of
> transaction properties needed by ApacheDS to comply with various
> specification requirements. For example all LDAP operations must be atomic
> and must preserve consistency. Furthermore, one can easily debate the need
> for isolation where any operation does not see or get impacted by the
> partial changes of another concurrent operation.
>
> We started discussing these ACID properties and ways in which they could be
> implemented. Isolation is very interesting now thanks to directory
> versioning and the change log. We can easily implement isolation now. When a
> relatively lengthy operation like a search is being conducted, it should not
> see modifications within scope that occur after the search began. The search
> operation in the example can be replaced with any other operation minus all
> unbind, some extended, and all abandon requests.
>   
Atomicity and Isolation are both complex to guarantee in a LDAP server.

If we think about Atomicity, for instance, even if we can guarantee it 
somehow for somesimple operation like Modify, Add or Delete, for the 
ModDN operation is not that simple. We have to guarantee that all the 
potential renames are done - or reverted - as a whole. As this operation 
can impact a big part of the server, and could take several seconds 
(minutes, hours, dependening on the number of entries), this is 
obviously not trivial. However, moddn operation aren't the most frequent 
one. Regarding the most simpler operation (add, delete and modify), I 
think we should implement some kind of "transaction" in the backend : 
the modified entry has to be tagged as 'under modification' until the 
backend has updated correctly the modification (or rollbacked it). Then 
we can untag the entry, and it's available back. How the CL can come 
into play here is to be discussed. IMHO, the CL and this 'transaction' 
will work hand to hand at some point.

Regarding isolation, it's a bit more difficult, as a search can already 
have sent back some results which could be change by another modify 
operation. This is especially the case for a ModDN operation.
> The change log, not only tracks each change, but it allows the directory
> server to inject the "revisions" attribute into entries. The revisions
> attribute is multi-valued and lists all the revisions of changes which
> altered the entry. For the search example, we can conduct the search
> operation while locking it down to a revision. 
That does not work for deleted entries, obviously ...
>  This is best implemented by
> conditionally filtering out or injecting candidates with revisions greater
> than the revision at which the search operation started. Let's call the
> revision when search started S. So entries in the server which posses
> revision numbers greater than S need further evaluation. We have to evaluate
> if the filter matches these entries with revisions > S when their state was
> at revision S. This may require some reverse patching and re-evaluation of
> the filter on the patched entry in state S.  This is not so bad because
> there usually are not that many changes taking place at the same time on the
> same entry: meaning the number of LDIF's to patch on an entry to evaluate in
> it's former state at S will be small. This way we effectively lock down the
> search to a specific revision, giving the search operation what appears to
> be a snapshot of the DIT. The search results will not be impacted by any
> concurrent changes.
>   
Well, I don't think this is the best approach. In any case, a Ldap 
Search is considered as a dirty read. We have no way to lock down the 
modification on the read entries. So the user has _no_ guarantee that 
the entry he gets back will be valid. Usually, it doesn't matter, 
because the ratio of read vs writes on a LDAP server is just so big that 
we consider we don't have modifications. So we can simply return the 
latest revision, whatever it is. Anyway, there is another aspect we have 
to consider : once the user gets his entry, and before he uses it, 
before potentially send it back as a modification to the server, the 
very same entry can have been modified in between. As we don't lock 
entries, we can't protect the users from such a case.

<digression>
We have to remember here that we are not dealing with bank accounts and 
balances, or nuclear plants. Most of the case, we are using a LDAP 
server to handle identities. They rarely change, or when they do, it's 
because the person owning this identity is changing his own identity - 
thus limitating the odds that he is using it at the same time -. 
Usually, if we think about authorizations, which are subject to way more 
changes that identities, we can't consider it as a continuous flow of 
modification.

In other words, creations or deletion of entries might be frequent, 
modification should be quite rare.

I have gathered some stats from some of my clients, and the ration 
changes/reads is like 1/5000... I would be interested to get more 
numbers here !

Last, not least, I consider that if the ratio goes up to something like 
1/100, then it's time to consider using a transactional system, namely, 
a RDBMS.
</digression>

So far, I'm not saying that it's wrong to think about using a MVCC 
system, but I'm a bit sceptic about the gain in term of isolation (the I 
in ACID) it will offer in our case.

Let's discuss this anyway, it's interesting !

-- 
--
cordialement, regards,
Emmanuel Lécharny
www.iktek.com
directory.apache.org