ignite-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Alex Plehanov <plehanov.a...@gmail.com>
Subject Re: Thin client: transactions support
Date Wed, 27 Mar 2019 10:36:54 GMT
Vladimir,

About current tx: ok, then we don't need tx() method in the interface at
all (the same cached transaction info user can store by himself).

About decoupling transactions from threads on the server side: for now, we
can start with thread-per-connection approach (we only can support one
active transaction per connection, see below, so we need one additional
dedicated thread for each connection with active transaction), and later
change server-side internals to process client transactions in any server
thread (not dedicated to this connection). This change will not affect the
thin client protocol, it only affects the server side.
In any case, we can't support concurrent transactions per connection on
the client side without fundamental changes to the current protocol (cache
operation doesn't bound to transaction or thread and the server doesn't
know which thread on the client side do this cache operation). In my
opinion, if a user wants to use concurrent transactions, he must use
different connections from a connection pool.

About semantics of suspend/resume on the client-side: it's absolutely
different than server-side semantics (we don't need to do suspend/resume to
pass transaction between threads on the client-side), but can't be
implemented efficiently without implemented suspend/resume on server-side.

Can anyone give me permissions to create IEP on Apache wiki?

ср, 27 мар. 2019 г. в 11:59, Vladimir Ozerov <vozerov@gridgain.com>:

> Hi Alex,
>
> My comments was only about the protocol. Getting current info about
> transaction should be handled by the client itself. It is not protocl's
> concern. Same about other APIs and behavior in case another transaction is
> attempted from the same thread.
>
> Putting protocol aside, transaction support is complicated matter. I would
> propose to route through IEP and wide community discussion. We need to
> review API and semantics very carefully, taking SUSPEND/RESUME in count.
> Also I do not see how we support client transactions efficiently without
> decoupling transactions from threads on the server side first. Because
> without it you will need a dedicated server thread for every client's
> transaction which is slow and may even crash the server.
>
> Vladimir.
>
> On Wed, Mar 27, 2019 at 11:44 AM Alex Plehanov <plehanov.alex@gmail.com>
> wrote:
>
> > Vladimir, what if we want to get current transaction info (tx() method)?
> >
> > Does close() method mapped to TX_END(rollback)?
> >
> > For example, this code:
> >
> > try(tx = txStart()) {
> >     tx.commit();
> > }
> >
> > Will produce:
> > TX_START
> > TX_END(commit)
> > TX_END(rollback)
> >
> > Am I understand you right?
> >
> > About xid. There is yet another proposal. Use some unique per connection
> id
> > (integer, simple counter) for identifying the transaction on
> > commit/rollback message. The client gets this id from the server with
> > transaction info and sends it back to the server when trying to
> > commit/rollback transaction. This id is not shown to users. But also we
> can
> > pass from server to client real transaction id (xid) with transaction
> info
> > for diagnostic purposes.
> >
> > And one more question: what should we do if the client starts a new
> > transaction without ending the old one? Should we end the old transaction
> > implicitly (rollback) or throw an exception to the client? In my opinion,
> > the first option is better. For example, if we got a previously used
> > connection from the connection pool, we should not worry about any
> > uncompleted transaction started by the previous user of this connection.
> >
> > ср, 27 мар. 2019 г. в 11:02, Vladimir Ozerov <vozerov@gridgain.com>:
> >
> > > As far as SUSPEND/RESUME/SAVEPOINT - we do not support them yet, and
> > adding
> > > them in future should not conflict with simple START/END
> infrastructure.
> > >
> > > On Wed, Mar 27, 2019 at 11:00 AM Vladimir Ozerov <vozerov@gridgain.com
> >
> > > wrote:
> > >
> > > > Hi Alex,
> > > >
> > > > I am not sure we need 5 commands. Wouldn't it be enough to have only
> > two?
> > > >
> > > > START - accepts optional parameters, returns transaction info
> > > > END - provides commit flag, returns void
> > > >
> > > > Vladimir.
> > > >
> > > > On Wed, Mar 27, 2019 at 8:26 AM Alex Plehanov <
> plehanov.alex@gmail.com
> > >
> > > > wrote:
> > > >
> > > >> Sergey, yes, the close is something like silent rollback. But we can
> > > >> also implement this on the client side, just using rollback and
> > ignoring
> > > >> errors in the response.
> > > >>
> > > >> ср, 27 мар. 2019 г. в 00:04, Sergey Kozlov <skozlov@gridgain.com>:
> > > >>
> > > >> > Nikolay
> > > >> >
> > > >> > Am I correctly understand you points:
> > > >> >
> > > >> >    - close: rollback
> > > >> >    - commit, close: do nothing
> > > >> >    - rollback, close: do what? (I suppose nothing)
> > > >> >
> > > >> > Also you assume that after commit/rollback we may need to free
> some
> > > >> > resources on server node(s)or just do on client started TX?
> > > >> >
> > > >> >
> > > >> >
> > > >> > On Tue, Mar 26, 2019 at 10:41 PM Alex Plehanov <
> > > plehanov.alex@gmail.com
> > > >> >
> > > >> > wrote:
> > > >> >
> > > >> > > Sergey, we have the close() method in the thick client,
it's
> > > behavior
> > > >> is
> > > >> > > slightly different than rollback() method (it should rollback
if
> > the
> > > >> > > transaction is not committed and do nothing if the transaction
> is
> > > >> already
> > > >> > > committed). I think we should support try-with-resource
> semantics
> > in
> > > >> the
> > > >> > > thin client and OP_TX_CLOSE will be useful here.
> > > >> > >
> > > >> > > Nikolay, suspend/resume didn't work yet for pessimistic
> > > transactions.
> > > >> > Also,
> > > >> > > the main goal of suspend/resume operations is to support
> > transaction
> > > >> > > passing between threads. In the thin client, the transaction
is
> > > bound
> > > >> to
> > > >> > > the client connection, not client thread. I think passing
a
> > > >> transaction
> > > >> > > between different client connections is not a very useful
case.
> > > >> > >
> > > >> > > вт, 26 мар. 2019 г. в 22:17, Nikolay Izhikov <
> nizhikov@apache.org
> > >:
> > > >> > >
> > > >> > > > Hello, Alex.
> > > >> > > >
> > > >> > > > We also have suspend and resume operations.
> > > >> > > > I think we should support them
> > > >> > > >
> > > >> > > > вт, 26 марта 2019 г., 22:07 Sergey Kozlov <
> skozlov@gridgain.com
> > >:
> > > >> > > >
> > > >> > > > > Hi
> > > >> > > > >
> > > >> > > > > Looks like I missed something but why we need
OP_TX_CLOSE
> > > >> operation?
> > > >> > > > >
> > > >> > > > > Also I suggest to reserve a code for SAVEPOINT
operation
> which
> > > >> very
> > > >> > > > useful
> > > >> > > > > to understand where transaction has been rolled
back
> > > >> > > > >
> > > >> > > > > On Tue, Mar 26, 2019 at 6:07 PM Alex Plehanov
<
> > > >> > plehanov.alex@gmail.com
> > > >> > > >
> > > >> > > > > wrote:
> > > >> > > > >
> > > >> > > > > > Hello Igniters!
> > > >> > > > > >
> > > >> > > > > > I want to pick up the ticket IGNITE-7369
and add
> > transactions
> > > >> > support
> > > >> > > > to
> > > >> > > > > > our thin client implementation.
> > > >> > > > > > I've looked at our current implementation
and have some
> > > >> proposals
> > > >> > to
> > > >> > > > > > support transactions:
> > > >> > > > > >
> > > >> > > > > > Add new operations to thin client protocol:
> > > >> > > > > >
> > > >> > > > > >     OP_TX_GET, 4000, Get current transaction
for client
> > > >> connection
> > > >> > > > > >     OP_TX_START, 4001, Start a new transaction
> > > >> > > > > >     OP_TX_COMMIT, 4002, Commit transaction
> > > >> > > > > >     OP_TX_ROLLBACK, 4003, Rollback transaction
> > > >> > > > > >     OP_TX_CLOSE, 4004, Close transaction
> > > >> > > > > >
> > > >> > > > > > From the client side (java) new interfaces
will be added:
> > > >> > > > > >
> > > >> > > > > > public interface ClientTransactions {
> > > >> > > > > >     public ClientTransaction txStart();
> > > >> > > > > >     public ClientTransaction
> txStart(TransactionConcurrency
> > > >> > > > concurrency,
> > > >> > > > > > TransactionIsolation isolation);
> > > >> > > > > >     public ClientTransaction
> txStart(TransactionConcurrency
> > > >> > > > concurrency,
> > > >> > > > > > TransactionIsolation isolation, long timeout,
int txSize);
> > > >> > > > > >     public ClientTransaction tx(); // Get
current
> connection
> > > >> > > > transaction
> > > >> > > > > >     public ClientTransactions withLabel(String
lb);
> > > >> > > > > > }
> > > >> > > > > >
> > > >> > > > > > public interface ClientTransaction extends
AutoCloseable {
> > > >> > > > > >     public IgniteUuid xid(); // Do we need
it?
> > > >> > > > > >     public TransactionIsolation isolation();
> > > >> > > > > >     public TransactionConcurrency concurrency();
> > > >> > > > > >     public long timeout();
> > > >> > > > > >     public String label();
> > > >> > > > > >
> > > >> > > > > >     public void commit();
> > > >> > > > > >     public void rollback();
> > > >> > > > > >     public void close();
> > > >> > > > > > }
> > > >> > > > > >
> > > >> > > > > > From the server side, I think as a first
step (while
> > > >> transactions
> > > >> > > > > > suspend/resume is not fully implemented)
we can use the
> same
> > > >> > approach
> > > >> > > > as
> > > >> > > > > > for JDBC: add a new worker to each ClientRequestHandler
> and
> > > >> process
> > > >> > > > > > requests by this worker if the transaction
is started
> > > >> explicitly.
> > > >> > > > > > ClientRequestHandler is bound to client connection,
so
> there
> > > >> will
> > > >> > be
> > > >> > > > 1:1
> > > >> > > > > > relation between client connection and thread,
which
> process
> > > >> > > operations
> > > >> > > > > in
> > > >> > > > > > a transaction.
> > > >> > > > > >
> > > >> > > > > > Also, there is a couple of issues I want
to discuss:
> > > >> > > > > >
> > > >> > > > > > We have overloaded method txStart with a
different set of
> > > >> > arguments.
> > > >> > > > Some
> > > >> > > > > > of the arguments may be missing. To pass
arguments with
> > > >> OP_TX_START
> > > >> > > > > > operation we have the next options:
> > > >> > > > > >  * Serialize full set of arguments and use
some value for
> > > >> missing
> > > >> > > > > > arguments. For example -1 for int/long types
and null for
> > > string
> > > >> > > type.
> > > >> > > > We
> > > >> > > > > > can't use 0 for int/long types since 0 it's
a valid value
> > for
> > > >> > > > > concurrency,
> > > >> > > > > > isolation and timeout arguments.
> > > >> > > > > >  * Serialize arguments as a collection of
property-value
> > pairs
> > > >> > (like
> > > >> > > > it's
> > > >> > > > > > implemented now for CacheConfiguration).
In this case only
> > > >> > explicitly
> > > >> > > > > > provided arguments will be serialized.
> > > >> > > > > > Which way is better? The simplest solution
is to use the
> > first
> > > >> > option
> > > >> > > > > and I
> > > >> > > > > > want to use it if there were no objections.
> > > >> > > > > >
> > > >> > > > > > Do we need transaction id (xid) on the client
side?
> > > >> > > > > > If yes, we can pass xid along with OP_TX_COMMIT,
> > > OP_TX_ROLLBACK,
> > > >> > > > > > OP_TX_CLOSE operations back to the server
and do
> additional
> > > >> check
> > > >> > on
> > > >> > > > the
> > > >> > > > > > server side (current transaction id for connection
==
> > > >> transaction
> > > >> > id
> > > >> > > > > passed
> > > >> > > > > > from client side). This, perhaps, will protect
clients
> > against
> > > >> some
> > > >> > > > > errors
> > > >> > > > > > (for example when client try to commit outdated
> > transaction).
> > > >> But
> > > >> > > > > > currently, we don't have data type IgniteUuid
in thin
> client
> > > >> > > protocol.
> > > >> > > > Do
> > > >> > > > > > we need to add it too?
> > > >> > > > > > Also, we can pass xid as a string just to
inform the
> client
> > > and
> > > >> do
> > > >> > > not
> > > >> > > > > pass
> > > >> > > > > > it back to the server with commit/rollback
operation.
> > > >> > > > > > Or not to pass xid at all (.NET thick client
works this
> way
> > as
> > > >> far
> > > >> > > as I
> > > >> > > > > > know).
> > > >> > > > > >
> > > >> > > > > > What do you think?
> > > >> > > > > >
> > > >> > > > > > ср, 7 мар. 2018 г. в 16:22, Vladimir
Ozerov <
> > > >> vozerov@gridgain.com
> > > >> > >:
> > > >> > > > > >
> > > >> > > > > > > We already have transactions support
in JDBC driver in
> TX
> > > SQL
> > > >> > > branch
> > > >> > > > > > > (ignite-4191). Currently it is implemented
through
> > separate
> > > >> > thread,
> > > >> > > > > which
> > > >> > > > > > > is not that efficient. Ideally we need
to finish
> > decoupling
> > > >> > > > > transactions
> > > >> > > > > > > from threads. But alternatively we can
change the logic
> on
> > > >> how we
> > > >> > > > > assign
> > > >> > > > > > > thread ID to specific transaction and
"impersonate" thin
> > > >> client
> > > >> > > > worker
> > > >> > > > > > > threads when serving requests from multiple
users.
> > > >> > > > > > >
> > > >> > > > > > >
> > > >> > > > > > >
> > > >> > > > > > > On Tue, Mar 6, 2018 at 10:01 PM, Denis
Magda <
> > > >> dmagda@apache.org>
> > > >> > > > > wrote:
> > > >> > > > > > >
> > > >> > > > > > > > Here is an original discussion
with a reference to the
> > > JIRA
> > > >> > > ticket:
> > > >> > > > > > > > http://apache-ignite-developers.2346864.n4.nabble.
> > > >> > > > > > > >
> > > com/Re-Transaction-operations-using-the-Ignite-Thin-Client-
> > > >> > > > > > > > Protocol-td25914.html
> > > >> > > > > > > >
> > > >> > > > > > > > --
> > > >> > > > > > > > Denis
> > > >> > > > > > > >
> > > >> > > > > > > > On Tue, Mar 6, 2018 at 9:18 AM,
Dmitriy Setrakyan <
> > > >> > > > > > dsetrakyan@apache.org
> > > >> > > > > > > >
> > > >> > > > > > > > wrote:
> > > >> > > > > > > >
> > > >> > > > > > > > > Hi Dmitriy. I don't think
we have a design proposal
> > for
> > > >> > > > transaction
> > > >> > > > > > > > support
> > > >> > > > > > > > > in thin clients. Do you mind
taking this initiative
> > and
> > > >> > > creating
> > > >> > > > an
> > > >> > > > > > IEP
> > > >> > > > > > > > on
> > > >> > > > > > > > > Wiki?
> > > >> > > > > > > > >
> > > >> > > > > > > > > D.
> > > >> > > > > > > > >
> > > >> > > > > > > > > On Tue, Mar 6, 2018 at 8:46
AM, Dmitriy Govorukhin <
> > > >> > > > > > > > > dmitriy.govorukhin@gmail.com>
wrote:
> > > >> > > > > > > > >
> > > >> > > > > > > > > > Hi, Igniters.
> > > >> > > > > > > > > >
> > > >> > > > > > > > > > I've seen a lot of discussions
about thin client
> and
> > > >> binary
> > > >> > > > > > protocol,
> > > >> > > > > > > > > but I
> > > >> > > > > > > > > > did not hear anything
about transactions support.
> Do
> > > we
> > > >> > have
> > > >> > > > some
> > > >> > > > > > > draft
> > > >> > > > > > > > > for
> > > >> > > > > > > > > > this purpose?
> > > >> > > > > > > > > >
> > > >> > > > > > > > > > As I understand we have
several problems:
> > > >> > > > > > > > > >
> > > >> > > > > > > > > >    - thread and transaction
have hard related (we
> > use
> > > >> > > > > thread-local
> > > >> > > > > > > > > variable
> > > >> > > > > > > > > >    and thread name)
> > > >> > > > > > > > > >    - we can process only
one transaction at the
> same
> > > >> time
> > > >> > in
> > > >> > > > one
> > > >> > > > > > > thread
> > > >> > > > > > > > > (it
> > > >> > > > > > > > > >    mean we need hold
thread per client. If connect
> > 100
> > > >> thin
> > > >> > > > > clients
> > > >> > > > > > > to
> > > >> > > > > > > > 1
> > > >> > > > > > > > > >    server node, then
need to hold 100 thread on
> the
> > > >> server
> > > >> > > > side)
> > > >> > > > > > > > > >
> > > >> > > > > > > > > > Let's discuss how we
can implement transactions
> for
> > > the
> > > >> > thin
> > > >> > > > > > client.
> > > >> > > > > > > > > >
> > > >> > > > > > > > >
> > > >> > > > > > > >
> > > >> > > > > > >
> > > >> > > > > >
> > > >> > > > >
> > > >> > > > >
> > > >> > > > > --
> > > >> > > > > Sergey Kozlov
> > > >> > > > > GridGain Systems
> > > >> > > > > www.gridgain.com
> > > >> > > > >
> > > >> > > >
> > > >> > >
> > > >> >
> > > >> >
> > > >> > --
> > > >> > Sergey Kozlov
> > > >> > GridGain Systems
> > > >> > www.gridgain.com
> > > >> >
> > > >>
> > > >
> > >
> >
>

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