10:37 a.m.
Some users have expressed the need to have some sort of forEach operation
that is performed where the Consumer is called while holding the lock for
the given key and subsequently released after the Consumer operation
completes.
Due to the nature of how streams work with retries and performing the
operation on the primary owner, this works out quite well with forEach to
be done in an efficient way.
The problem is that this only really works well with non tx and pessimistic
tx. This obviously leaves out optimistic tx, which at first I was a little
worried about. But after thinking about it more, this prelocking and
optimistic tx don't really fit that well together anyways. So I am thinking
whenever this operation is performed it would throw an exception not
letting the user use this feature in optimistic transactions.
Another question is what does the API for this look like. I was debating
between 3 options myself:
1. AdvancedCache.forEachWithLock(BiConsumer<Cache, CacheEntry<K, V>>
consumer)
This require the least amount of changes, however the user can't customize
certain parameters that CacheStream currently provides (listed below - big
one being filterKeys).
2. CacheStream.forEachWithLock(BiConsumer<Cache, CacheEntry<K, V>> consumer)
This method would only be allowed to be invoked on the Stream if no other
intermediate operations were invoked, otherwise an exception would be
thrown. This still gives us access to all of the CacheStream methods that
aren't on the Stream interface (ie. sequentialDistribution,
parallelDistribution, parallel, sequential, filterKeys, filterKeySegments,
distributedBatchSize, disableRehashAware, timeout).
3. LockedStream<CacheEntry<K, V>> AdvancedCache.lockedStream()
This requires the most changes, however the API would be the most explicit.
In this case the LockedStream would only have the methods on it that are
able to be invoked as noted above and forEach.
I personally feel that #3 might be the cleanest, but obviously requires
adding more classes. Let me know what you guys think and if you think the
optimistic exclusion is acceptable.
Thanks,
- Will
11:16 a.m.
I'm leaning towards option 1.
Are you thinking about also allowing the consumer to modify the entry,
like JCache's EntryProcessors? For a consumer that can only modify the
current entry, we could even "emulate" locking in an optimistic cache
by catching the WriteSkewException and running the consumer again.
I wouldn't allow this to be mixed with other operations in a stream,
because then you may have to run filters/mappers/sorting while holding
the lock as well.
Cheers
Dan
On Tue, Mar 21, 2017 at 5:37 PM, William Burns <mudokonman(a)gmail.com> wrote:
12:31 p.m.
On Tue, Mar 21, 2017 at 12:17 PM Dan Berindei <dan.berindei(a)gmail.com>
wrote:
I'm leaning towards option 1.
This is actually what I have implemented right now. The big problem I find
is the lack of configuring the operation. For example I know that users
will require being able to pass in a set of keys. The more I think about it
a Predicate would also be preferred. This could be handled by having
another override of the method that takes this Predicate or Set of keys. My
main concern was whether we thought users may want to tweak other settings
such as timeout, parallel or sequential operation per node or cluster and
also batch size possibly to reduce chances of more than once operations.
This ends up exploding the overrides, which is why I was hoping to use
something like the Stream interface to handle each as a different method
invocation. This is what brought me to think of #2, and subsequently #3
after I thought of excluding all those methods.
That is one possible case (updating cache entry). To be honest the user
could want it for other reasons that I am not privy to.
The catching of WriteSkewException is interesting, but I don't think it is
acceptable. The problem is that the value could change between which could
in turn change what the user wants to write or operation they perform and
there is no way to inform them.
That was my idea as well and why I was preventing those in all three APIs.
Although as I mentioned above allowing filter should be fine too. It also
makes the process of locking and unlocking the entry quite cumbersome if we
allowed other operations.
8:49 a.m.
On Mon, Mar 27, 2017 at 9:02 PM Galder Zamarreño <galder(a)redhat.com> wrote:
No the primary owner would run the operation per entry. I was thinking we
would have 2 levels of filtering in my proposal above.
We would have the first one which is using filterKeys on the CacheStream
method. This requires serializing keys sent to each node (only primary
owned keys are sent). While it has the cost of serialization it makes up
for by having constant time lookups (no iterating memory/stores) for the
keys as it creates a stream using Cache.get to populate it.
The second was to support the filter method on the Stream API which would
allow for a Predicate so you don't have to serialize keys. In this case you
wouldn't want to include keys in this Predicate as all keys would be
serialized to all nodes and then you still have to iterate and check the
entire data container/store.
You could actually do both as well. So if you only want a subset of known
keys where their values match a Predicate this can be done too.
cache.lockedStream().filterKeys(keys).filter(predicate).forEach();
9:55 a.m.
Hi Will,
I'm confused about the premise; when you state
" the Consumer is called while holding the lock for the given key and
subsequently released after the Consumer operation completes."
What are the transaction boundaries?
I see two options, please correct me so I understand:
A)
transaction.begin();
cache.lockedStream().filterKeys(keys).filter(predicate).forEach(
someOperation );
transaction.commit(); <-- release all locks
B)
cache.lockedStream().filterKeys(keys).filter(predicate).forEach(
transaction.begin();
//to stuff on entry
/transaction.commit(); <-- release this single entry's lock
);
I think it's important to clarify this as I suspect that #A is not
implementable within reasonable guarantees, while in the case of B# I
see no use for optimistic locking *from a user's perspective*.
Also: what kind of guarantees do you expect about having the operation
being applied on some entry which is strictly *still* a subset of the
keys as defined by the filter predicates?
I take it you'd want to acquire the locks during the filtering process?
That would require the move the transaction boundary to the scenario
A# which seems undesirable.
Technically if I were to need something like this I guess I'd expect
to have a user experience akin to B# but have Infinispan essentially
use optimistic locking (and auto-skip) on entries which are mutated
and fall out of the filter predicate during the lock attempt.
Essentially I suspect that we'd not want to implement different
versions of this depending on the transaction mode, but figure out the
use case and implement a one and only transaction mode which suites
such use cases. So for example we'd simply not offer a mode which
requires to copy the whole grid into the current TX context.
Thanks,
Sanne
On 28 March 2017 at 14:49, William Burns <mudokonman(a)gmail.com> wrote:
10:44 a.m.
On Tue, Mar 28, 2017 at 11:24 AM Sanne Grinovero <sanne(a)infinispan.org>
wrote:
The user code doesn't really affect this, it is done internally by
Infinispan. We would acquire the lock on the user's behalf in the stream
operation then call the user's accept method. Then after they return we
would unlock the lock. The way I have it implemented at the moment in my
PoC (which could change) is I don't even start a transaction. It just locks
the key and then if the user were to invoke an operation that requires a
transaction it adds the lock owner to their tx context at that point so
they have ownership of the key.
Also to note that a commit or rollback in the Consumer currently doesn't
release the lock on the key. Although this could be discussed to be
possibly changed.
Exactly my thoughts regarding optimistic. I don't think #A is even feasible
given constraints of having a distributed transaction like this.
That is a good question. I hadn't thought about the details but what I had
implemented was we have to first read the entry, lock the key, reread the
entry (in case of concurrent update) and then finally call their Predicate.
Another reason the filterKeys is much more efficient :) Note this read,
locking and reread is done even without a Predicate supplied. And actually
I can also optimize filterKeys to not do the initial read since we already
have the key.
This was never the intent and in my follow up emails I came to what seems
like the same conclusion that basically this can't be done with the user
controlling the transaction and it doesn't really make sense in an
optimistic transaction (since you are already at that node, you are already
doing everything pessimistically).
11:44 a.m.
On Tue, Mar 28, 2017 at 6:44 PM, William Burns <mudokonman(a)gmail.com> wrote:
With a transactional cache I was assuming you manage the transaction
yourself... if the user has to call
transactionManager.begin()/commit()/rollback() anyway, why not use a
regular stream?
Totally agree that #A can't work with any kind of transaction configuration.
As to optimistic locking, I would like having "feature parity" between
pessimistic and optimistic caches as much as possible, but I agree
locking eagerly and retrying the consumer on WriteSkewException are a
bit too different to fit under the same API.
Would this be more efficient than the consumer reading the key with
FORCE_WRITE_LOCK and deciding what to do based on the value?
Even local caches can use optimistic locking :)
12:54 p.m.
On Tue, Mar 28, 2017 at 12:52 PM Dan Berindei <dan.berindei(a)gmail.com>
wrote:
The user would be using an implicit transaction or explicit. Like I said
this is up for discussion. The main reason I am staying away from managing
the transaction is that the user can mess with the transaction as well
which would possibly release the lock. It is much simpler if all I am doing
is managing the lock. And if the user doesn't require a transaction in
Infinispan we didn't waste time starting one and releasing one.
Yeah I would definitely like to have it as well, but I just can't see how
it fits in. This is despite the implementation detail that it is quite
difficult to get it working currently :D
The problem with this is you have to go remote to lock the key, return the
value then do something for every key (not to mention pulling those keys
using an iterator). Very costly! The main benefit of the stream with lock
is that you are performing everything on the primary owner of the data with
the lock already acquired. The only piece sent remotely is the consumer and
some internal classes, very light weight and you have all the benefits of
data locality.
Yes I know :) I was just framing it in the notion of remote. If anyone can
think of a nice way of using this with optimistic transactions I would all
be for it. But optimistic transactions just doesn't make any sense to me
when you are locking a key eagerly for someone to do something with it
(definition of pessimistic transaction).
3:04 a.m.
On Tue, Mar 28, 2017 at 8:54 PM, William Burns <mudokonman(a)gmail.com> wrote:
Every write to a transactional cache will start an implicit
transaction, so I don't think we really have the option of not
starting a transaction.
I meant using cache.withFlags(FORCE_WRITE_LOCK).get(key) in the
Consumer itself, and assuming it runs on the primary owner.
Maybe I didn't explain properly...
I meant in optimistic caches we wouldn't acquire the lock at all. The
stream would just start an optimistic transaction, run the consumer,
and try to commit. If the prepare fails because of a
WriteSkewException, start from the beginning.
8:24 a.m.
On Wed, Mar 29, 2017 at 4:05 AM Dan Berindei <dan.berindei(a)gmail.com> wrote:
Yes a write to it will, but the stream stuff doesn't write to it, it just
locks. So if the user doesn't have to actually write to the cache, they may
want to just update a 3rd party system for example.
Ah yes you could for pessimistic, but not optimistic or nontx :( In this
case it would be very similar.
This assumes the user will be updating the cache, I am not sure we can
always assume that (in regards to starting a transaction).
But this is a good point (I misunderstood earlier), it should be pretty
easy to just catch WriteSkewException and retry the given Consumer with the
new entry (if it exists still).
10:37 a.m.
On Wed, Mar 29, 2017 at 4:24 PM, William Burns <mudokonman(a)gmail.com> wrote:
Yeah, my main use case was "processing" all/most of the entries in the
cache in some way. I must confess I didn't even think about using
Infinispan locks to synchronize access to external systems, I was
thinking only what would happen when the Consumer would access
multiple caches :)
OTOH a read-only transaction doesn't have to be very expensive either.
We already have an optimization in PessimisticLockingInterceptor to
skip sending the LockControlCommand to backup owners [1]. This applies
only for LockControlCommands generated implicitly by writes, but we
could modify TxDistributionInterceptor.visitLockControlCommand() to
skip the RPC for explicit AdvancedCache.lock() calls as well. If there
are no writes and no remote locks, the prepare and commit should be
local as well.
There will still be some overhead from interacting with the
transaction manager, and I can understand wanting to avoid that. I'm
not yet convinced it's worth writing a new API and potentially
confusing lock release behaviour because of it...
[1]:
https://github.com/infinispan/infinispan/blob/9fb8752d847a85e168d785237f1...
11:51 a.m.
On 03/21/2017 04:37 PM, William Burns wrote:
Seconding Dan's question - is that intended to be able to modify the
entry? In my opinion, sending a function that will work on the
ReadWriteEntryView directly to the node is the only reasonable request.
I wouldn't like to see blocking operations in there.
How exactly reading streams interacts with transactions? Does it wrap
read entries into context? This would be a scalability issue.
I agree that "locking" should not be exposed with optimistic transactions.
With pessimistic transactions, how do you expect to handle locking
order? For regular operations, user is responsible for setting up some
locking order in order to not get a deadlock. With pessimistic
transaction, it's the cache itself who will order the calls. Also, if
you lock anything that is read, you just end up locking everything (or,
getting a deadlock). If you don't it's the same as issuing the lock and
reading again (to check the locked value) - but you'd do that internally
anyway. Therefore, I don't feel well about pessimistic transactions neither.
For both options, I don't like Cache being passed around. You should
modify the CacheEntry (or some kind of view) directly.
Radim
--
Radim Vansa <rvansa(a)redhat.com>
JBoss Performance Team
12:42 p.m.
On Tue, Mar 21, 2017 at 12:53 PM Radim Vansa <rvansa(a)redhat.com> wrote:
Hrmm the user can use the FunctionalMap interface for this then it seems? I
wonder if this should just be the going in API. I will need to discuss with
Galder the semantics of the evalAll/evalMany methods.
It doesn't wrap read entries into the context for that exact reason. It
does however use existing entries in the context to override ones in
memory/store.
I agree that "locking" should not be exposed with optimistic transactions.
Yeah I can't find a good way to do this really and it seems to be opposite
of what optimistic transactions are.
The lock is done per key only for each invocation. There is no ordering as
only one is obtained at a time before it goes to the next. If the user then
acquires a lock for another key while in the Consumer this could cause a
deadlock if the inverse occurs on a different thread/node, but this is on
the user. It is the same as it is today really, except we do the read lock
for them before invoking their Consumer.
I don't know for sure if that is sufficient for the user. Sometimes they
may modify another Cache given the value in this one for example, which
they could access from the CacheManager of that Cache. Maybe Tristan knows
more about some use cases.
4:51 a.m.
On 03/21/2017 06:50 PM, William Burns wrote:
Don't look into the implementation; I think Galder has focused more on
the API side than having optimal implementation. IMO there's no reason
evalAll should load all the entries into memory in non-transactional mode.
Uuh, so you end up with a copy of the cache in single invocation
context, without any means to flush it. I think that we need add
InvocationContext.current().forget(key) API (throwing exception if the
entry was modified) or something like that, even for the regular
streams. Maybe an override for filter methods, too, because you want to
pass a nice predicate, but you can't just forget all filtered out entries.
In pessimistic mode, you should not release a lock before the end of the
transaction.
Rather than guessing what could the user need, the Consumer could be CDI
enabled.
--
Radim Vansa <rvansa(a)redhat.com>
JBoss Performance Team
8:41 a.m.
On Wed, Mar 22, 2017 at 5:51 AM Radim Vansa <rvansa(a)redhat.com> wrote:
I agree that it shouldn't, but there is no guarantee this will be ready any
time soon.
Maybe I worded it poorly. Streams don't wrap any entries at all. All it
does is read from the current context if there any, it then reads from the
data container (skipping entries it read from the context) and then finally
reading from the store if present.
Although the more I think about it using Stream with lock this may be a non
issue, read below.
I agree. I didn't discuss finer details, but what I have right now doesn't
work with an ongoing pessimistic transaction. And to be honest I am not
sure if this can work with an ongoing transaction. And even if it did it
would be horrendously horrendously slow since each remote node is
performing the Consumer but it would need to read from the originators
context (or have it copied over). Unless someone with more transaction
knowledge knows of a better way I personally feel using stream with lock
should run in its own dedicated transaction.
I am not the biggest CDI fan, especially since the benefit of this is you
can use lambdas (that automatically become Serializable) and make very
concise code. But others can chime in here too.
9:33 a.m.
On Wed, Mar 22, 2017 at 11:51 AM, Radim Vansa <rvansa(a)redhat.com> wrote:
I'm pretty sure we do need to load all the entries in order to provide
REPEATABLE_READ isolation.
I think Will said he *doesn't* want to wrap the entries read by the consumer :)
IMO there's no "good" way to provide repeatable read isolation for a
transaction that reads all the keys in the cache, so this API should
create a separate transaction for each entry. I wouldn't try to make
the consumers see the current transaction's modifications if started
from a transaction either, I'd throw an exception if started from a
transaction instead.
Ok, the name forEachWithLock doesn't really fit with optimistic
locking, but I think with a more neutral name it could work for
optimistic caches as well.
Exactly. Each consumer needs to have its own transaction, otherwise
the transaction's lockedKeys collection would have to grow to include
all the keys in the cache.
If the user actually needs to work with more than one entry at a time,
I think it would be much cleaner for him to use regular forEach() and
start an explicit transaction in the consumer.
8:42 a.m.
On Tue, Mar 28, 2017 at 9:27 AM Galder Zamarreño <galder(a)redhat.com> wrote:
It actually is very similar to the distributed iterator. Basically it
performs distributedBatchSize number of consumer calls then tells the
originator it finished those keys sending them back and continues on. This
way at most you can have distributedBatchSize * numNodes worth of more than
once calls per state transfer. In practice this will be significantly lower
though as you would only have a subset and you have to lose complete
ownership of said key.
While keys are sent across the wire, values are never returned with
forEach. And key references are released when a segment is completed so it
should have a subset of keys in memory.
3:43 p.m.
On Tue, Mar 21, 2017 at 4:28 PM Tristan Tarrant <ttarrant(a)redhat.com> wrote:
I would probably add a new interface that defines the common methods
between CacheStream and LockedStream (or whatever we call it). Then the
actual LockedStream class should be a self contained class. Thinking about
it more it shouldn't require more than a couple hundred extra lines of
code. The hard part was just getting the locking to work with non tx and
pessimistic tx, which is common to any impl.
Also just to note you would still get an UnsupportedOperationException when
running in optimistic tx when invoking the method.
2924
days inactive
2932
days old
infinispan-dev@lists.jboss.org
22 comments
6 participants
participants (6)
-
Dan Berindei -
Galder Zamarreño -
Radim Vansa -
Sanne Grinovero -
Tristan Tarrant -
William Burns