[infinispan-dev] Feedback and requests on clustered and remote listeners
mudokonman at gmail.com
Fri Sep 19 11:09:51 EDT 2014
Comments regarding embedded usage are inline. I am not quite sure on
the hot rod client ones.
On Thu, Sep 18, 2014 at 12:24 PM, Emmanuel Bernard
<emmanuel at hibernate.org> wrote:
> Hi all,
> I have had a good exchange on how someone would use clustered / remote listeners to do custom continuous query features.
> I have a few questions and requests to make this fully and easily doable
> ## Value as bytes or as objects
> Assuming a Hot Rod based usage and protobuf as the serialization layer. What are KeyValueFilter and Converter seeing?
> I assume today the bytes are unmarshalled and the Java object is provided to these interfaces.
> In a protobuf based storage, does that mean that the user must create the Java objects out of a protobuf compiler and deploy these classes in the classpath of each server node?
> Alternatively, could we pass the raw protobuf data to the KeyValueFilter and Converter? They could read the relevant properties at no deserialization cost and with lss problems related to the classloader.
> ## Synced listeners
> In a transactional clustered listener marked as sync. Does the transaction commits and then waits for the relevant clustered listeners to proceed before returning the hand to the Tx client? Or is there something else going on?
It commits the transaction and then notifies the listeners. The
listener notification is done while still holding all appropriate
locks for the given entry to guarantee proper ordering.
> ## oldValue and newValue
> I understand why the oldValue was not provided in the initial work. It requires to send more data across the network and at least double the number of values unmarshalled.
> But for continuous queries, being able to compare the old and the new value is critical to reduce the number of events sent to the listener.
> Imagine the following use case. A listener exposes the average age for a certain type of customer. You would implement it the following way.
> 1. Add a KeyValueFilter that
> - upon creation, filter out the customers of the wrong type
> - upon update, keep customers that
> - *were* of the right time but no longer are
> - were not of the right type but now now *are*
> - remains of the right type and whose age has changed
> - upon deletion, keep customers that *were* of the right type
> 2. Converter
> In the converter, one could send the whole customer but it would be more efficient to only send the age of the customer as well as wether it is added to or removed from the matching customers
> - upon creation, you send the customer age and mark it as addition
> - upon deletion, you send the customer age and mark it as deletion
> - upon update
> - if the customer was of the right type but no longer is, send the age as well as a deletion flag
> - if the customer was not of the right type but now is, send the age as well as an addition flag
> - if the customer age has changed, send the difference with a modification flag
> 3. The listener then needs to keep the total sum of all ages as well as the total number of customers of the right type. Based on the sent events, it can adjust these two counters.
> That requires us to be able to provide the old and new value to the KeyValueFilter and the Converter interface as well as the type of event (creation, update, deletion).
I agree the oldValue is required for most efficient usage. From the
oldValue though it seems you can infer what operation it is. Create
has null oldValue and delete has null newValue I would think.
This also came up here
and I am debating if this interface should be separate or just an
extension from KeyValueFilter etc.
The thing is the new interface is mostly beneficial only to clustered
listeners since non cluster listeners get both the pre and post event
which makes the old value accessible. I may have to just try to write
it up and see how it goes unless anyone has any suggestions.
> If you keep the existing interfaces and their data, the data send and the memory consumed becomes much much bigger. I leave it as an exercise but I think you need to:
> - send *all* remove and update events regardless of the value (essentially no KeyValueFilter)
> - in the listener, keep a list of *all* matching keys so that you know if a new event is about a data that was already matching your criteria or not and act accordingly.
> BTW, you need the old and new value even if your listener returns actual matching results instead of an aggregation. More or less for the same reasons.
> Continuous query is about the most important use case for remote and clustered listeners and I think we should address it properly and as efficiently as possible. Adding continuous query to Infinispan will then “simply” be a matter of agreeing on the query syntax and implement the predicates as smartly as possible.
> With the use case I describe, I think the best approach is to merge the KVF and Converter into a single Listener like interface that is able to send or silence an event payload. But that’s guestimate.
> Because oldValue / newValue implies an unmarshalling overhead we might want to make it an annotation based flag on the class that is executed on each node (somewhat similar to the settings hosted on @Listener).
We actually have an interface that combines the 2 interfaces, it is
called KeyValueFilterConverter. It was added to more efficiently
perform indexless querying using entry retriever. This interface is
not supported for cluster listeners at this time though.
> ## includeCurrentState and very narrow filtering
> The existing approach is fine (send a create event notif for all existing keys and queue changes in the mean time) as long as the listener plans to consume most of these events.
> But in case of a big data grid, with a lot of passivated entries, the cost would become non negligible.
The filter and converter are applied while doing the current state so
it should be performant in that case. Also to note while the current
state operation is ongoing any new notifications are enqueued until
the current state is applied. These new events will not cause
blocking as you mentioned earlier with sync since they are immediately
enqueued. The queueing may be something we have to add blocking
though possibly to prevent memory exhaustion in the case when the
initial iteration is extremely slow and there are a lot of updates
during that period. The code currently has code to release queued
events by segment as the segments are completed, I have thought about
also releasing events by key instead which should relieve a lot of
possible memory usage.
> An alternative approach is to first do a query matching the elements the listener is interested in and queue up the events until the query is fully processed. Can a listener access a cache and do a query? Should we offer such option in a more packaged way?
The provided filter be doing this already.
But maybe more info on what you are proposing. Either way it seems we
have to have the listener installed before we can run the query so we
can properly tell what events should be raised in the event of
concurrent events while the query is running.
> For a listener that is only interested in keys whose value city contains Springfield, Virginia, the gain would be massive.
> ## Remote listener and non Java HR clients
> Does the API of non Java HR clients support the enlistements of listeners and attach registered keyValueFilter / Converter? Or is that planned? Just curious.
> infinispan-dev mailing list
> infinispan-dev at lists.jboss.org
More information about the infinispan-dev