ignite-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Vladimir Ozerov <voze...@gridgain.com>
Subject Re: Continuous queries and duplicates
Date Wed, 09 Jan 2019 20:10:57 GMT
Hi,

MVCC caches have the same ordering guarantees as non-MVCC caches, i.e. two
subsequent updates on a single key will be delivered in proper order. There
is no guarantees  Order of updates on two subsequent transactions affecting
the same partition may be guaranteed with current implementation (though. I
am not sure), but even if it is so, I am not aware that this was ever our
design goal. Most likely, this is an implementation artifact which may be
changed in future. Cache experts are needed to clarify this.

As far as MVCC, data anomalies are still possible in current
implementation, because we didn't rework initial query handling in the
first iteration, because technically this is not so simple as we thought.
Once snapshot is obtained, query over that snapshot will return a data set
consistent at some point in time. But the problem is that there is a time
frame between snapshot acquisition and listener installation (or vice
versa), what leads to either duplicates or lost entries. Some multi-step
listener installation will be required here. We haven't designed it yet.

Vladimir.



On Mon, Dec 24, 2018 at 10:06 PM Denis Magda <dmagda@apache.org> wrote:

> >
> > In my case, values are immutable - I never change them, I just add new
> > entry for newer versions. Does it mean that I won't have any duplicates
> > between the initial query and listener entries when using continuous
> > queries on caches supporting MVCC?
>
>
> I'm afraid there still might be a race. Val, Vladimir, other Ignite
> experts, please confirm.
>
> After reading the related thread (
> >
> >
> http://apache-ignite-developers.2346864.n4.nabble.com/Continuous-queries-and-MVCC-td33972.html
> > )
> > I'm now concerned about the ordering. My case assumes that there are
> groups
> > of entries which belong to a business aggregate object and I would like
> to
> > make sure that if I commit two records in two serial transactions then I
> > have notifications in the same order. Those entries will have different
> > keys so based on what you said ("we'd better to leave things as is and
> > guarantee only per-key ordering"), it would seem that the order is not
> > guaranteed. But do you think it would possible to guarantee order when
> > those entries share the same affinity key and they belong to the same
> > partition?
>
>
> The order should be the same for key-value transactions. Vladimir, could
> you clear out MVCC based behavior?
>
> --
> Denis
>
> On Mon, Dec 17, 2018 at 9:55 AM Piotr Romański <piotr.romanski@gmail.com>
> wrote:
>
> > Hi all, sorry for answering so late.
> >
> > I would like to use SqlQuery because I can leverage indexes there.
> >
> > As it was already mentioned earlier, the partition update counter is
> > exposed through CacheQueryEntryEvent. Initially, I thought that the
> > partition update counter is something what's persisted together with the
> > data but I'm guessing now that this is only a part of the notification
> > mechanism.
> >
> > I imagined that I would be able to implement my own deduplicaton by
> having
> > 3 stages on the client side: 1. Keep processing initial query results,
> > store their keys in memory, 2. When initial query is over, then process
> > listener entries but before that check if they have been already
> delivered
> > in the first stage, 3. When we are sure that we are already processing
> > notifications for commits executed after initial query was done, then we
> > can process listener entries without any additional checks (so our key
> set
> > from stage 1 can be removed from memory). The problem is that I have no
> way
> > to say that I can move from stage 2 to 3. Another problem is that we need
> > to stash listener entries while still processing initial query results
> > causing an excessive memory pressure on our client.
> >
> > In my case, values are immutable - I never change them, I just add new
> > entry for newer versions. Does it mean that I won't have any duplicates
> > between the initial query and listener entries when using continuous
> > queries on caches supporting MVCC?
> >
> > After reading the related thread (
> >
> >
> http://apache-ignite-developers.2346864.n4.nabble.com/Continuous-queries-and-MVCC-td33972.html
> > )
> > I'm now concerned about the ordering. My case assumes that there are
> groups
> > of entries which belong to a business aggregate object and I would like
> to
> > make sure that if I commit two records in two serial transactions then I
> > have notifications in the same order. Those entries will have different
> > keys so based on what you said ("we'd better to leave things as is and
> > guarantee only per-key ordering"), it would seem that the order is not
> > guaranteed. But do you think it would possible to guarantee order when
> > those entries share the same affinity key and they belong to the same
> > partition?
> >
> > Piotr
> >
> > pt., 14 gru 2018, 19:31: Denis Magda <dmagda@apache.org> napisał(a):
> >
> > > Vladimir,
> > >
> > > Thanks for referring to the MVCC and Continuous Queries discussion, I
> > knew
> > > that saw us discussing a solution of the duplication problem. Let me
> copy
> > > and paste it in here for others:
> > >
> > > 2) *Initial query*. We implemented it so that user can get some initial
> > > > data snapshot and then start receiving events. Without MVCC we have
> no
> > > > guarantees of visibility. E.g. if key is updated from V1 to V2, it is
> > > > possible to see V2 in initial query and in event. With MVCC it is now
> > > > technically possible to query data on certain snapshot and then
> receive
> > > > only events happened after this snapshot. So that we never see V2
> > twice.
> > > > Do
> > > > you think we this feature will be interesting for our users?
> > >
> > >
> > > Am I right that this would be a generic solution - whether you use Scan
> > or
> > > SQL query as an initial one? Have we planned it for the transactional
> SQL
> > > GA or it's out of scope for now?
> > >
> > > --
> > > Denis
> > >
> > > On Thu, Dec 13, 2018 at 12:40 PM Vladimir Ozerov <vozerov@gridgain.com
> >
> > > wrote:
> > >
> > > > [1]
> > > >
> > > >
> > >
> >
> http://apache-ignite-developers.2346864.n4.nabble.com/Continuous-queries-and-MVCC-td33972.html
> > > >
> > > > On Thu, Dec 13, 2018 at 11:38 PM Vladimir Ozerov <
> vozerov@gridgain.com
> > >
> > > > wrote:
> > > >
> > > > > Denis,
> > > > >
> > > > > Not really. They are used to ensure that ordering of notifications
> is
> > > > > consistent with ordering of updates, so that when a key K is
> updated
> > to
> > > > V1,
> > > > > then V2, then V3, you never observe V1 -> V3 -> V2. It also
solves
> > > > > duplicate notification problem in case of node failures, when the
> > same
> > > > > update is delivered twice.
> > > > >
> > > > > However, partition counters are unable to solve duplicates problem
> in
> > > > > general. Essentially, the question is how to get consistent view
on
> > > some
> > > > > data plus all notifications which happened afterwards. There are
> only
> > > two
> > > > > ways to achieve this - either lock entries during initial query,
or
> > > take
> > > > a
> > > > > kind of consistent data snapshot. The former was never implemented
> in
> > > > > Ignite - our Scan and SQL queries do not user locking. The latter
> is
> > > > > achievable in theory with MVCC. I raised that question earlier [1]
> > (see
> > > > > p.2), and we came to conclusion that it might be a good feature for
> > the
> > > > > product. It is not implemented that way for MVCC now, but most
> > probably
> > > > is
> > > > > not extraordinary difficult to implement.
> > > > >
> > > > > Vladimir.
> > > > >
> > > > > [1]
> > > > >
> > > >
> > >
> >
> http://apache-ignite-developers.2346864.n4.nabble.com/Continuous-queries-and-MVCC-td33972.html#a33998
> > > > >
> > > > > On Thu, Dec 13, 2018 at 11:17 PM Denis Magda <dmagda@apache.org>
> > > wrote:
> > > > >
> > > > >> Vladimir,
> > > > >>
> > > > >> The partition counter is supposed to be used internally to solve
> the
> > > > >> duplication issue. Does it sound like a right approach then?
> > > > >>
> > > > >> What would be an approach for SQL queries? Not sure the partition
> > > > counter
> > > > >> is applicable.
> > > > >>
> > > > >> --
> > > > >> Denis
> > > > >>
> > > > >> On Thu, Dec 13, 2018 at 11:16 AM Vladimir Ozerov <
> > > vozerov@gridgain.com>
> > > > >> wrote:
> > > > >>
> > > > >> > Partition counter is internal implemenattion detail, which
has
> no
> > > > >> sensible
> > > > >> > meaning to end users. It should not be exposed through public
> API.
> > > > >> >
> > > > >> > On Thu, Dec 13, 2018 at 10:14 PM Denis Magda <dmagda@apache.org
> >
> > > > wrote:
> > > > >> >
> > > > >> > > Hello Piotr,
> > > > >> > >
> > > > >> > > That's a known problem and I thought a JIRA ticket
already
> > exists.
> > > > >> > However,
> > > > >> > > failed to locate it. The ticket for the improvement
should be
> > > > created
> > > > >> as
> > > > >> > a
> > > > >> > > result of this conversation.
> > > > >> > >
> > > > >> > > Speaking of an initial query type, I would differentiate
from
> > > > >> ScanQueries
> > > > >> > > and SqlQueries. For the former, it sounds reasonable
to apply
> > the
> > > > >> > > partitionCounter logic. As for the latter, Vladimir
Ozerov
> will
> > it
> > > > be
> > > > >> > > addressed as part of MVCC/Transactional SQL activities?
> > > > >> > >
> > > > >> > > Btw, Piotr what's your initial query type?
> > > > >> > >
> > > > >> > > --
> > > > >> > > Denis
> > > > >> > >
> > > > >> > > On Thu, Dec 13, 2018 at 3:28 AM Piotr Romański <
> > > > >> piotr.romanski@gmail.com
> > > > >> > >
> > > > >> > > wrote:
> > > > >> > >
> > > > >> > > > Hi, as suggested by Ilya here:
> > > > >> > > >
> > > > >> > > >
> > > > >> > >
> > > > >> >
> > > > >>
> > > >
> > >
> >
> http://apache-ignite-users.70518.x6.nabble.com/Continuous-queries-and-duplicates-td25314.html
> > > > >> > > > I'm resending it to the developers list.
> > > > >> > > >
> > > > >> > > > From that thread we know that there might be duplicates
> > between
> > > > >> initial
> > > > >> > > > query results and listener entries received as
part of
> > > continuous
> > > > >> > query.
> > > > >> > > > That means that users need to manually dedupe
data.
> > > > >> > > >
> > > > >> > > > In my opinion the manual deduplication in some
use cases may
> > > lead
> > > > to
> > > > >> > > > possible memory problems on the client side. In
order to
> > remove
> > > > >> > > duplicated
> > > > >> > > > notifications which we are receiving in the local
listener,
> we
> > > > need
> > > > >> to
> > > > >> > > keep
> > > > >> > > > all initial query results in memory (or at least
their
> unique
> > > > ids).
> > > > >> > > > Unfortunately, there is no way (is there?) to
find a point
> in
> > > time
> > > > >> when
> > > > >> > > we
> > > > >> > > > can be sure that no dups will arrive anymore.
That would
> mean
> > > that
> > > > >> we
> > > > >> > > need
> > > > >> > > > to keep that data indefinitely and use it every
time a new
> > > > >> notification
> > > > >> > > > arrives. In case of multiple continuous queries
run from a
> > > single
> > > > >> JVM,
> > > > >> > > this
> > > > >> > > > might eventually become a memory or performance
problem. I
> can
> > > see
> > > > >> the
> > > > >> > > > following possible improvements to Ignite:
> > > > >> > > >
> > > > >> > > > 1. The deduplication between initial query and
incoming
> > > > notification
> > > > >> > > could
> > > > >> > > > be done fully in Ignite. As far as I know there
is already
> the
> > > > >> > > > updateCounter and partition id for all the objects
so it
> could
> > > be
> > > > >> used
> > > > >> > > > internally.
> > > > >> > > >
> > > > >> > > > 2. Add a guarantee that notifications arriving
in the local
> > > > listener
> > > > >> > > after
> > > > >> > > > query() method returns are not duplicates. This
kind of
> > > > >> functionality
> > > > >> > > would
> > > > >> > > > require a specific synchronization inside Ignite.
It would
> > also
> > > > mean
> > > > >> > that
> > > > >> > > > the query() method cannot return before all potential
> > duplicates
> > > > are
> > > > >> > > > processed by a local listener what looks wrong.
> > > > >> > > >
> > > > >> > > > 3. Notify users that starting from a given notification
they
> > can
> > > > be
> > > > >> > sure
> > > > >> > > > they will not receive any duplicates anymore.
This could be
> an
> > > > >> > additional
> > > > >> > > > boolean flag in the CacheQueryEntryEvent.
> > > > >> > > >
> > > > >> > > > 4. CacheQueryEntryEvent already exposes the
> > > > partitionUpdateCounter.
> > > > >> > > > Unfortunately we don't have this information for
initial
> query
> > > > >> results.
> > > > >> > > If
> > > > >> > > > we had, a client could manually deduplicate notifications
> and
> > > get
> > > > >> rid
> > > > >> > of
> > > > >> > > > initial query results for a given partition after
newer
> > > > >> notifications
> > > > >> > > > arrive. Also it would be very convenient to expose
partition
> > id
> > > as
> > > > >> well
> > > > >> > > but
> > > > >> > > > now we can figure it out using the affinity service.
The
> > > > assumption
> > > > >> > here
> > > > >> > > is
> > > > >> > > > that notifications are ordered by partitionUpdateCounter
(is
> > it
> > > > >> true?).
> > > > >> > > >
> > > > >> > > > Please correct me if I'm missing anything.
> > > > >> > > >
> > > > >> > > > What do you think?
> > > > >> > > >
> > > > >> > > > Piotr
> > > > >> > > >
> > > > >> > >
> > > > >> >
> > > > >>
> > > > >
> > > >
> > >
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message