I'm also not sure but for some reason I feel that we may need some event
loop to be implemented if we want to re-write it to async capabilities.
The reason I'm telling it is because I feel that in some situations we may
trigger a very long call stack.
I.e.:
Promise -> Promise -> Promise -> ....
I guess that could be in a situation when the next part of the execution
depends on the previous part of the execution. For example,
g.V().has("hello", "world").barrier(50).limit(5)
So, let's assume the next things about the execution of the above query:
- It is executed in JanusGraph with batch query enabled (graph provider
specific, but it's easier for me to focus on a concrete implementation)
- There is no necessary index for that property (again, graph provider
specific)
- There are only 4 vertices with such property
- We have 50 million vertices in total
If the above facts are true then the query will be executed like the
following:
1) get first (or next) 50 vertices
2) filter out unmatched vertices
3) if the limit is not reached 5 then process the next vertices by starting
from step "1" again. Otherwise, return data.
So, in fact, with the above scenario we will traverse all 50 million
vertices in the graph which will result in about 1 million chain calls for
promises based implementation. That will probably result in
StackOverflowException.
With synchronous code we don't have these problems because we don't call a
new function recursively each time we need to retrieve part of the data.
We can overcome the above issue by implementing some kind of event loop
where we put all the results and then a single thread running that loop
will call necessary async functions.
If so, we will never have a long chain of calls. The only thing in such
architecture is that it will require all those functions to be non-blocking
and thus all providers will probably need to implement their async
functionality
(doesn't matter if it is real async or "fake" async behind a thread pool).
That's just something on top of my mind, but I could be wrong and could
miss some points. So, maybe we don't need any event loop, I just didn't
investigate it enough yet.
Oleksandr
On Fri, Jul 29, 2022 at 1:09 PM Oleksandr Porunov <
[email protected]> wrote:
> Hi Divij,
>
> Thanks for joining the conversation!
>
> Basically the `promise` step is for remote queries execution only which is
> seen from the `promis` method "throw new IllegalStateException("Only
> traversals created using withRemote() can be used in an async way");".
> Under the hood Gremlin Server will be doing sync execution still even if
> the graph provider can support async query execution.
> The Gremlin Server will need to allocate 200 threads if we are trying to
> execute 200 parallel queries.
> g.V().has("name", "hello").out("world") - this query usually will first
> try to find vertex id for `hello` name and block the thread until it is
> processed. After that it will fire another query to the storage backend to
> retrieve all adjacent vertices following the "world" edge and block the
> executing thread until we receive a response from the database. After that
> it will return data back to the client.
> Instead of that I wanted to propose a different execution strategy.
> Let's fire the first query and instead of waiting for the response let's
> say when the response is available, please execute another step - which is
> get all adjacent vertices following the "world" edge. At that step we are
> doing the same, we are firing a database query and instead of waiting for
> the response we say, when the response is available, please execute another
> step (which is kind of a final step, which simply returns data to the
> client by executing a necessary function).
> I guess something like this would help to not allocate unnecessary threads
> which simply wait for the response from the underlying storage backends.
>
> Oleksandr
>
> On Fri, Jul 29, 2022 at 8:34 AM Divij Vaidya <[email protected]>
> wrote:
>
>> Hey folks
>>
>> Interesting discussion!
>>
>> I am a bit confused though since I believe we already have async
>> execution implemented in TinkerPop Java client. Let me try to clarify and
>> please let me know if I missed something.
>>
>> Java client uses a small number of websocket connections to multiplex
>> multiple queries to the server. You can think of it as a pipe established
>> to the server on which we could send messages belonging to different
>> queries. On the server, these messages are queued until one of the
>> execution threads can pick it up. Once a request is picked for execution,
>> the results are returned in a pipelines/streaming manner i.e. the server
>> calculates a batch of results (size of batch is configurable per query),
>> and sends the results as messages on the same WebSocket channel. On the
>> client size, these results are stored in a queue until the application
>> thread consumes them uses an iterator. This model of execution *does not
>> block the application thread* and hence, provides async capabilities.
>>
>> A sample code to achieve this would be as follows:
>>
>> ```
>> final Cluster cluster = Cluster.build("localhost")
>> .port(8182)
>>
>> .maxInProcessPerConnection(32)
>>
>> .maxSimultaneousUsagePerConnection(32)
>>
>> .serializer(Serializers.GRAPHBINARY_V1D0)
>> .create();
>>
>> try {
>> final GraphTraversalSource g =
>> traversal().withRemote(DriverRemoteConnection.using(cluster));
>> CompletableFuture<List<Object>> result = g.V().has("name",
>> "pumba").out("friendOf").id().promise(Traversal::toList);
>>
>> // do some application layer stuff
>> // ...
>> // ...
>> // ...
>>
>> List<Object> verticesWithNamePumba = result.join();
>> System.out.println(verticesWithNamePumba);
>> } finally {
>> cluster.close();
>> }
>> ```
>>
>> Note that, in the above example, the thread executing the above code is
>> not blocked until we call "result.join()".
>>
>> Does this address the use that Oleksandr brought up at the beginning of
>> this thread?
>>
>> --
>> Divij Vaidya
>>
>>
>>
>> On Fri, Jul 29, 2022 at 4:05 AM Oleksandr Porunov <
>> [email protected]> wrote:
>>
>>> Hmm, that's interesting! Thank you Joshua for the idea!
>>> So, I guess the general idea here could be:
>>> we can start small and start implementing async functionality for some
>>> parts instead of implement async functionality for everything
>>> straightaway.
>>>
>>> Oleksandr
>>>
>>> On Fri, Jul 29, 2022, 00:38 Joshua Shinavier <[email protected]> wrote:
>>>
>>> > Well, the wrapper I mentioned before did not require a full rewrite of
>>> > TinkerPop :-) Rather, it provided async interfaces for vertices and
>>> edges,
>>> > on which operations like subgraph and shortest paths queries were
>>> evaluated
>>> > in an asynchronous fashion (using a special language, as it happened,
>>> but
>>> > limited Gremlin queries would have been an option). So I think a basic
>>> > async API might be a useful starting point even if it doesn't go very
>>> deep.
>>> >
>>> > Josh
>>> >
>>> >
>>> > On Thu, Jul 28, 2022 at 4:21 PM Oleksandr Porunov <
>>> > [email protected]> wrote:
>>> >
>>> >> Hi Joshua and Pieter,
>>> >>
>>> >> Thank you for joining the conversation!
>>> >>
>>> >> I didn't actually look into the implementation details yet but quickly
>>> >> checking Traversal.java code I think Pieter is right here.
>>> >> For some reason I thought we could simply wrap synchronous method in
>>> >> asynchronous, basically something like:
>>> >>
>>> >> // the method which should be implemented by a graph provider
>>> >>
>>> >> Future<E> executeAsync(Callable<E> func);
>>> >>
>>> >> public default Future<E> asyncNext(){
>>> >> return executeAsync(this::next);
>>> >> }
>>> >>
>>> >> but checking that code I think I was wrong about it. Different steps
>>> may
>>> >> execute different logic (i.e. different underlying storage queries)
>>> for
>>> >> different graph providers.
>>> >> Thus, wrapping only terminal steps into async functions won't solve
>>> the
>>> >> problem most likely.
>>> >>
>>> >> I guess it will require re-writing or extending all steps to be able
>>> to
>>> >> pass an async state instead of a sync state.
>>> >>
>>> >> I'm not familiar enough with the TinkerPop code yet to claim that, so
>>> >> probably I could be wrong.
>>> >> I will need to research it a bit more to find out but I think that
>>> Pieter
>>> >> is most likely right about a massive re-write.
>>> >>
>>> >> Nevertheless, even if that requires massive re-write, I would be
>>> eager to
>>> >> start the ball rolling.
>>> >> I think we either need to try to implement async execution in
>>> TinkerPop 3
>>> >> or start making some concrete decisions regarding TinkerPop 4.
>>> >>
>>> >> I see Marko A. Rodriguez started to work on RxJava back in 2019 here
>>> >>
>>> https://github.com/apache/tinkerpop/tree/4.0-dev/java/machine/processor/rxjava/src/main/java/org/apache/tinkerpop/machine/processor/rxjava
>>> >>
>>> >> but the process didn't go as far as I understand. I guess it would be
>>> >> good to know if we want to completely rewrite TinkerPop in version 4
>>> or not.
>>> >>
>>> >> If we want to completely rewrite TinkerPop in version 4 then I assume
>>> it
>>> >> may take quite some time to do so. In this case I would be more
>>> likely to
>>> >> say that it's better to implement async functionality in TinkerPop 3
>>> even
>>> >> if it requires rewriting all steps.
>>> >>
>>> >> In case TinkerPop 4 is a redevelopment with breaking changes but
>>> without
>>> >> starting to rewrite the whole functionality then I guess we could try
>>> to
>>> >> work on TinkerPop 4 by introducing async functionality and maybe
>>> applying
>>> >> more breaking changes in places where it's better to re-work some
>>> parts.
>>> >>
>>> >> Best regards,
>>> >> Oleksandr
>>> >>
>>> >>
>>> >> On Thu, Jul 28, 2022 at 7:47 PM pieter gmail <[email protected]
>>> >
>>> >> wrote:
>>> >>
>>> >>> Hi,
>>> >>>
>>> >>> Does this not imply a massive rewrite of TinkerPop? In particular the
>>> >>> iterator chaining pattern of steps should follow a reactive style of
>>> >>> coding?
>>> >>>
>>> >>> Cheers
>>> >>> Pieter
>>> >>>
>>> >>>
>>> >>> On Thu, 2022-07-28 at 15:18 +0100, Oleksandr Porunov wrote:
>>> >>> > I'm interested in adding async capabilities to TinkerPop.
>>> >>> >
>>> >>> > There were many discussions about async capabilities for TinkerPop
>>> >>> > but
>>> >>> > there was no clear consensus on how and when it should be
>>> developed.
>>> >>> >
>>> >>> > The benefit for async capabilities is that the user calling a query
>>> >>> > shouldn't need its thread to be blocked to simply wait for the
>>> result
>>> >>> > of
>>> >>> > the query execution. Instead of that a graph provider should take
>>> >>> > care
>>> >>> > about implementation of async queries execution.
>>> >>> > If that's the case then many graph providers will be able to
>>> optimize
>>> >>> > their
>>> >>> > execution of async queries by handling less resources for the query
>>> >>> > execution.
>>> >>> > As a real example of potential benefit we could get I would like to
>>> >>> > point
>>> >>> > on how JanusGraph executes CQL queries to process Gremlin queries.
>>> >>> > CQL result retrieval:
>>> >>> >
>>> >>>
>>> https://github.com/JanusGraph/janusgraph/blob/15a00b7938052274fe15cf26025168299a311224/janusgraph-cql/src/main/java/org/janusgraph/diskstorage/cql/function/slice/CQLSimpleSliceFunction.java#L45
>>> >>> >
>>> >>> > As seen from the code above, JanusGraph already leverages async
>>> >>> > functionality for CQL queries under the hood but JanusGraph is
>>> >>> > required to
>>> >>> > process those queries in synced manner, so what JanusGraph does -
>>> it
>>> >>> > simply
>>> >>> > blocks the whole executing thread until result is returned instead
>>> of
>>> >>> > using
>>> >>> > async execution.
>>> >>> >
>>> >>> > Of course, that's just a case when we can benefit from async
>>> >>> > execution
>>> >>> > because the underneath storage backend can process async queries.
>>> If
>>> >>> > a
>>> >>> > storage backend can't process async queries then we won't get any
>>> >>> > benefit
>>> >>> > from implementing a fake async executor.
>>> >>> >
>>> >>> > That said, I believe quite a few graph providers may benefit from
>>> >>> > having a
>>> >>> > possibility to execute queries in async fashion because they can
>>> >>> > optimize
>>> >>> > their resource utilization.
>>> >>> > I believe that we could have a feature flag for storage providers
>>> >>> > which
>>> >>> > want to implement async execution. Those who can't implement it or
>>> >>> > don't
>>> >>> > want to implement it may simply disable async capabilities which
>>> will
>>> >>> > result in throwing an exception anytime an async function is
>>> called.
>>> >>> > I
>>> >>> > think it should be fine because we already have some feature flags
>>> >>> > like
>>> >>> > that for graph providers. For example "Null Semantics" was added in
>>> >>> > TinkerPop 3.5.0 but `null` is not supported for all graph
>>> providers.
>>> >>> > Thus,
>>> >>> > a feature flag for Null Semantics exists like
>>> >>> > "g.getGraph().features().vertex().supportsNullPropertyValues()".
>>> >>> > I believe we can enable async in TinkerPop 3 by providing async as
>>> a
>>> >>> > feature flag and letting graph providers implement it at their
>>> will.
>>> >>> > Moreover if a graph provider wants to have async capabilities but
>>> >>> > their
>>> >>> > storage backends don't support async capabilities then it should be
>>> >>> > easy to
>>> >>> > hide async execution under an ExecutorService which mimics async
>>> >>> > execution.
>>> >>> > I believe we could do that for TinkerGraph so that users could
>>> >>> > experiment
>>> >>> > with async API at least. I believe we could simply have a default
>>> >>> > "async"
>>> >>> > function implementation for TinkerGraph which wraps all sync
>>> >>> > executions in
>>> >>> > a function and sends it to that ExecutorService (we can discuss
>>> which
>>> >>> > one).
>>> >>> > In such a case TinkerGraph will support async execution even
>>> without
>>> >>> > real
>>> >>> > async functionality. We could also potentially provide some
>>> >>> > configuration
>>> >>> > options to TinkerGraph to configure thread pool size, executor
>>> >>> > service
>>> >>> > implementation, etc.
>>> >>> >
>>> >>> > I didn't think about how it is better to implement those async
>>> >>> > capabilities
>>> >>> > for TinkerPop yet but I think reusing a similar approach like in
>>> >>> > Node.js
>>> >>> > which returns Promise when calling Terminal steps could be good.
>>> For
>>> >>> > example, we could have a method called `async` which accepts a
>>> >>> > termination
>>> >>> > step and returns a necessary Future object.
>>> >>> > I.e.:
>>> >>> > g.V(123).async(Traversal.next())
>>> >>> > g.V().async(Traversal.toList())
>>> >>> > g.E().async(Traversal.toSet())
>>> >>> > g.E().async(Traversal.iterate())
>>> >>> >
>>> >>> > I know that there were discussions about adding async functionality
>>> >>> > to
>>> >>> > TinkerPop 4 eventually, but I don't see strong reasons why we
>>> >>> > couldn't add
>>> >>> > async functionality to TinkerPop 3 with a feature flag.
>>> >>> > It would be really great to hear some thoughts and concerns about
>>> it.
>>> >>> >
>>> >>> > If there are no concerns, I'd like to develop a proposal for
>>> further
>>> >>> > discussion.
>>> >>> >
>>> >>> > Best regards,
>>> >>> > Oleksandr Porunov
>>> >>>
>>> >>>
>>>
>>
