Rick,
HA is totally transparent to any client of the embedded API after
setup. That means, if you have a number of Neo4j Server (or any other
thing like various bindings etc) the only difference is in the
instantiation of the GraphDatabaseService.
So, from the REST (or your own app) API you are not seeing HA at all.
It just looks like one DB running. To scale, you can set up one
endpoint to every HA slave in your cluster, and build your custom load
balancing (e.g. with AKKA etc) on top of that. See
http://docs.neo4j.org/html/milestone/#_high_availability for more
info.
Will be happy to help you give it a spin!
Where the REST layer might come into play is the upcoming sharding
effort - going to start working on that after 1.2, that is, pretty
soon :)
Cheers,
/peter neubauer
GTalk: neubauer.peter
Skype peter.neubauer
Phone +46 704 106975
LinkedIn http://www.linkedin.com/in/neubauer
Twitter http://twitter.com/peterneubauer
http://www.neo4j.org - Your high performance graph database.
http://www.thoughtmade.com - Scandinavia's coolest Bring-a-Thing party.
On Sat, Dec 4, 2010 at 6:32 PM, Rick Bullotta
rick.bullo...@burningskysoftware.com wrote:
Hi, Peter.
That makes sense. I was anticipating roughly 10X-100X difference. Good to
see we're near the lower end of that, though I'm not surprised. You guys
make fast stuff! ;-)
Another question: Is the REST API integral to the HA scenario or do you
envision that at some point the HA model be implemented with embedded
instances also? I'm just a bit unclear on how the HA/REST interaction (or
more specifically, the HA architecture) works. Any Wiki page I should look
at?
Thanks!
Rick
-Original Message-
From: user-boun...@lists.neo4j.org [mailto:user-boun...@lists.neo4j.org] On
Behalf Of Peter Neubauer
Sent: Saturday, December 04, 2010 12:17 PM
To: Neo4j user discussions
Subject: Re: [Neo4j] Performance benchmarks of REST API vs embedded Neo4J
Well,
we REST having the serialization overhead, plus the crossing of system
boundaries regarding caching, and the problem with transaction
handling, it is no doubt much more performant to use the embedded
Neo4j scenario when you want optimum performance and can control all
aspects of your system.
However, REST is great for integration and easy start, and access
outside Java. The third approach is a native driver protocol, maybe
built on Protobuffer, Thrift or others, that should in performance be
somewhere in between.
The Neo4j REST API is not on par with the native approach especially
regarding many (small) operations, we first need to enable better
server side support with either scripts being executed serverside
(dynamic via sending them or static via server extensions). Rexster is
simply exposing Jersey and JAXRS for Java classes, taking away some of
the burden of building your own server. We are right now debating on
how to best make server extensions both live near the Neo4j instance
and at the same time be isolated from the server, so the server can
perform without custom extension destroying things, much like stored
procedures in RDBMS. It looks like OSGi might perform good services
here ...
Anyway, benchmarks are hard to do since the range of functionality is
very different. In traversal scenarios though I think it is, and the
difference should not be too big, since most of the job is done on the
server side, the main overhead is the serialization of the request and
the results. Again, it depend on how your usage scenarios look like.
For small transactions, I have been running an insert of 1 nodes
via addNode() over REST with https://github.com/maxdemarzi/neography
in about 10sec, while the same takes less than 1sec on my machine in
embedded mode (IIRC) ... small transactions needing to hit disc.
Cheers,
/peter neubauer
GTalk: neubauer.peter
Skype peter.neubauer
Phone +46 704 106975
LinkedIn http://www.linkedin.com/in/neubauer
Twitter http://twitter.com/peterneubauer
http://www.neo4j.org - Your high performance graph database.
http://www.thoughtmade.com - Scandinavia's coolest Bring-a-Thing party.
On Sat, Dec 4, 2010 at 3:22 PM, Rick Bullotta
rick.bullo...@burningskysoftware.com wrote:
Has anyone done any rough performance comparisons of the two approaches?
I
have to think we're looking at 1 or 2 orders of magnitude difference, but
would like to know if there is any hard data yet.
___
Neo4j mailing list
User@lists.neo4j.org
https://lists.neo4j.org/mailman/listinfo/user
___
Neo4j mailing list
User@lists.neo4j.org
https://lists.neo4j.org/mailman/listinfo/user
___
Neo4j mailing list
User@lists.neo4j.org
https://lists.neo4j.org/mailman/listinfo/user
___
Neo4j
