jackrabbit-oak-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Thomas Mueller <muel...@adobe.com>
Subject Re: Observation
Date Thu, 07 Nov 2013 10:13:48 GMT

Sounds good to me! I was originally against adding a new API, but now that
I see all the problems with the JCR observation API, I think it a very
good idea.

One possible addition: as far as I know, for "remove" events, some
observation listeners would like to read the old data. I think Oak could
provide that data, at least in most cases. Would it make sense to add this
feature? Would it be hard to support it? (This would be just for Oak, not
Jackrabbit 2.x.)


On 11/7/13 2:49 AM, "Alexander Klimetschek" <aklimets@adobe.com> wrote:

>Hi everyone,
>having looked at and discussed observation recently (partly on our Adobe
>internal list with some of our application use cases, partly with Toby
>f2f), here are some things that are IMHO worthwile to look at. I hope I
>can contribute a bit, but before I get pulled away again, I wanted at
>least summarize and persist() my ideas:
>1) listeners should not get external events by default
>OAK-1121 helps, but applications must say they exclude external events,
>while it should be the other way around (only local by default) - this
>can arguably not change because of the jcr observation contract. The
>current implementation can still be optimized a bit, see my comment in
>the issue.
>2) filtering happens too late within listeners because jcr observation
>API is too simple (OAK-1133, "observation listener plus")
>The filter logic would be passed declaratively and allow more options as
>noted in OAK-1133:
>	addEventListener(listener, Filter.Path("/libs"),
>Filter.PropertyValue("jcr:content/sling:resourceType", "myValue"), ...)
>(just a rough sketch; the simplest solution could simply introduce a
>FilterEventListener interface that extends the existing EventListener and
>provides the filter in a getFilter() method; a simple instanceof check
>within oak-jcr could detect that; but a completely separate API might
>also make sense, see point 6 below)
>This could then be evaluated when the NodeStateDiff is calculated. What
>is good now already is that the base path (if specified in the jcr
>observation) is evaluated as soon as getting the diff roots. But many
>real use cases might want to have multiple paths (/apps OR /libs) or
>globbing (/content/site/*/jcr:content) - this could be easily and
>efficiently added here in one place (amongst those other filters). Having
>one or few base paths is in the oak case probably the best way to filter
>things and one should avoid listeners that register at root, and only
>filter on node types for example, which requires a lot more tree walking.
>In our app (CQ) we have 18 listeners that register for all events,
>because the JCR observation doesn't give them the right filter options,
>so they have to do it themselves. I expect them to be vastly improved
>with this new approach.
>Also, getting events for deleted nodes could be supported here, since
>there is full access to the old NodeState.
>The ideal solution could also be built backwards compatible: implement
>the same filtering with normal JCR API usage and automatically wrap the
>listener in another event listener that filters the "manual" way. This
>allows application code that works with both Jackrabbit 2 or Oak with
>minimal effort (would need some helper in jackrabbit-jcr-commons or so).
>Unless we want to backport this to Jackrabbit 2 as well.
>3) cluster events
>As agreed, external events can be the real scalability issue. It was also
>noted, that many cases do not need it at all (hence 1). But there are
>still cases that need it: in Sling for example, code deployment (bundles,
>JSP scripts, etc.) is based on the JCR repository, meaning when you put a
>bundle in a certain "install" folder in the JCR, all cluster instances
>need to pick it up. However, such deployment events are comparably  rare
>(except on a developer machine :)), so the throughput doesn't have to be
>high. OTOH large-scale content changes will (and should) be handled
>locally, thus not requiring external propagation.
>Now the question is if 2) with a filter-as-detailed-and-early-as-possible
>approach solves this already (having the right external/non-external flag
>on all listeners, plus detailed filters for the external listeners), or
>if we could make use of the fairly rare nature of those events and try to
>filter on the _source_ instance already and propagate those few events in
>some other, efficient way to other cluster instances. This might improve
>performance as the ChangeDispatcher#externalChange() check (polling all
>100ms) could then be removed completely. But if the filtering on the
>NodeStateDiff is fast (because listeners are specific), then maybe it is
>4) listener threads
>Currently in oak each event listener gets its own thread. In our app CQ
>we have about 150 listeners, so you end up introducing 150 threads. I am
>not sure if this is a good idea.
>Jackrabbit 2 had a single thread, which had the issue of one observation
>listener blocking all others.
>An obvious solution would be a thread pool (probably configurable number
>of threads). And on top of that, if the pool is full and no thread is
>free anymore, one could simply kill blocked handlers after a certain
>timeout (and optionally blacklist them).
>5) filtering in central thread?
>There was some discussion about the current filtering evaluation [0].
>Each listener has its own ChangeProcessor which in turn has its own
>ChangeDispatcher.Listener. This means filtering is not in a single,
>central thread, but happens in each listener's thread (they each have
>their own ChangeSet queue, which indeed might be different because the
>previous root NodeState to compare to might be different for each).
>I thought this was not optimal but it seems I was wrong - splitting that
>up and generating events in a central thread and then passing them in a
>queue to the listeners actually turned out to scale worse (in a quick
>test with many parallel listeners). See my patch at [1].
>I guess it is that reading asynchronously from the immutable NodeStates
>is more efficient than multiple blocking queues. Which speaks good for
>the underlying oak implementation :)
>[0] http://markmail.org/thread/533orsfr44wllvrx
>6) long-running session just for observation
>One problem with JCR observation is that you need one session open all
>the time "just for the observation". This is done to be able to run
>observation under the permissions of a certain user. But nobody in
>practice uses the session for anything else than reading data upon events
>(which mostly is done to filter only, see 2); when you need to write
>things in an event, best practice is to create a new session on demand an
>close it again. With Oak's refresh() policy this is even more important.
>Maybe a new API from 2) could skip the requirement of an open session.
>Registration would be based upon a session to handle the credentials
>easily, but that session could be closed without killing the listener.
>You would get back a special listener object that would need to be kept
>alive and referenced to be able to clean up listeners automatically on a
>finalize() if they forget to unregister.
>Or you keep the session, but mark it in a special way so that Oak can
>internally save resources. Or maybe the session in oak is already so
>light-weight that this is actually no problem at all. I am just
>mentioning this because of the many warnings I get from oak in the logs
>about sessions that are open very long (and don't call refresh() IIRC).
>7) move code to oak-jcr (minor)
>Quite a bit of the code in oak.plugins.observation is currently JCR
>eventing specific (EventGenerator for example), afaics this is better
>suited in oak-jcr. Although with 2) it might change a bit anyway.

View raw message