[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] [List Home]
Re: [ecf-dev] Problem with Service Endpoint matching when using different network names

Hi Scott,

that sounds and looks really promising! I'll need some time to process the information you gave and look at the code and then come back to you.

Have a nice weekend!

Bye Peter


Am 27.07.2018 um 03:54 schrieb Scott Lewis:
Hi Peter,

I couldn't resist.ÂÂ I've created a TCP socket-based distribution provider [1].ÂÂ Like we've been discussing, it's client->server only, and so clients cannot export services (and servers cannot import services), but unlike the other example impls using tcp that I'm aware of, it supports multiple remote calls via a single connection, and it supports multiple remote services exported via a single socket listener.ÂÂ It also supports the OSGi R7 async remote services and the other R7 features (intents, configurable timeout, etc).ÂÂ And it just uses ServerSocket and Socket classes, so there's a chance it will work in your target environment (if the environment allows non-known apps to make direct socket connections with an unknown...not JDBC, not http, etc protocol.

Of course it's not done and fully tested yet, but I did get signs of life this evening.

Scott

[1] https://github.com/ECF/tcpsocketdistributionprovider

On 7/26/2018 10:52 AM, Scott Lewis wrote:
Hi Peter,

On 7/26/2018 1:41 AM, Peter Hermsdorf wrote:

Hi Scott,


Am 21.07.2018 um 00:11 schrieb Scott Lewis:
When the mapping is done, who/what does the mapping?..and how is it done?  It seems to me that's the problematic case. Does it map both IP address and port, or just IP address? Do you know if it's using NAT or some other tech?

Are there other services on this net...e.g. a web server...that are working properly with the addressing properties that you are looking for?ÂÂ If so, how is that done?ÂÂ A reverse proxy or load balancing hw or ?
Typically mapping is only done for the ip addresses. Actually i can't tell exactly how it's done. Probably it's custom to every customer....
Other services that work that way are JDBC connections to oracle databases. They don't care if you reach them by hostname or the one ip address or the other ... ;)

Given the smiley, I'm not sure if you are joking but here's the rub:ÂÂ JDBC connections are point-to-point (strict client-server)...and so the addressing is relatively simple.ÂÂ Part of its simplicity is that it creates isolation between clients, but with DB connections that's generally what you want.

However, the ECF generic provider...and some of the others...provides a group model, where every process (clients and server) can both export and import remote services as opposed to server-export and client-import only.

Just to explain a little more:ÂÂ In a group model (i.e. ECF container), every process in the group has to

a) have a unique identity;
b) agree (membership) to use the same ID to refer to the same process.ÂÂ

So in the three group members case:

Serverid -> ecftcp://some.name:3333/server

Client1id -> 1

Client2id -> 2

When the two clients connect to this server, all three processes receive the IDs of the other two processes...i.e. Server gets 1, 2, Client1 gets serverid, 2; and Client2 gets serverid and 1.ÂÂÂ Note that if the Client1 serverid != Client2 serverid (i.e. the clients are on different networks) then it violates b above.ÂÂÂ This seems to be your situation with the generic provider.

All I'm saying is that the introduction of NAT, firewalls, proxies, VPNs, etc changes the addressing.ÂÂ This is less of a problem for client-server communication because there are only two processes 'aware' of each other instead of a group.

When I wrote the generic provider (originally > 14 years ago), the addressing introduced by NAT, VPN, etc wasn't nearly as prevalent. I *could* have introduced some additional connection/group join protocol to associate some separate/unique name (uuid, etc) with the server ip address, so that clients didn't use some.name:3333. However, at that time I didn't anticipate it would be necessary, and so I didn't do that...using the (guaranteed unique) some.name:3333 to both identify the process and client uses to connect to the server. In retrospect it would have been nice if I did, but OTOH given the complexity involved in doing it in the 'general case' I'm kind of glad I didn't :).ÂÂ

It's possible that a new/extended generic container could be created that had this additional protocol to have connected clients use a non-ip-based name for the server in the group.ÂÂ It would probably be necessary, however, to first understand what the name mapping was doing for a given network topology (i.e. your customer) at least if one was interested in keeping the 'group' nature (i.e. not be 'strict client-server' at the service level).ÂÂÂÂ

As we've been discussing, another option is to use a strict client-server topology rather than a group, and use or create a distribution provider based upon a strict client-server model.ÂÂ See below for more comments about this.

<stuff deleted>

1)Â currently we only use "strict" client->server setup, but: i could image use-cases where it could be useful if the server could import services from the clients to realize something like push information to the client from the server (without polling etc)Â - but that's probably another story...

Indeed it is :).

If having a strict client->server works for your services, then I would suggest you try the either the JaxRSRemoteService providers [1] which are based upon HttpService (jetty server usually).ÂÂ It still seems to me that you would need a reverse proxy like nginx to expose the same server to access via multiple IP addresses/networks, and I'm not sure if that's possible on your target network, but nginx is frequently used for that.

2) i don't think it's a good idea to switch to a http based communication - performance wise. i would like to stick with a binary transportation layer rather than have http protocol overhead (remember my kryo serialization implementation)... and we don't have a use case where other services/participants would benefit from a http based communication...
we have a jetty on server side, but it has nothing to deal with the osgi remote services - just provides some jax-rs rest services

Given that you already have a jetty server working in this topology, perhaps it would be worth it to give it a try with the JaxRSProvider...and see how the performance is for a test service.ÂÂ I understand the concern about performance with http...especially if you are sending lots of messages.ÂÂ However, as you know jetty/websockets, caching proxies, hw, etc., etc have improved the performance of http under many usage scenarios so maybe it will be less of a problem than you think.

Another thought:ÂÂ Once you were confident that a strict client-server model would get you want you want in terms of connection, you could create a simple websocket-or-regular-socket-based distribution provider based upon your Kyro serialization provider [1] or at least starting from that.ÂÂ With Photon I've tried to make it easier to create new distribution providers (more/more useful abstract classes).ÂÂ There are a other providers at [2] that you could model from (e.g. Chronicle, grcp, etc) or just use a simple socket connection based upon the trivial provider [3].

You can also combine multiple distribution providers if you need to (i.e. some services with JaxRS, others with a custom-socket-based distribution provider for others).

<stuff deleted>

3) switching to a different provider is an option, if there is "no" performance problem and this "connection" issue would be solved. additional infrastructure for translating/mapping/proxying is a problem and is at the end no real option.... from my point of view that's the job of the underlying tcp/ip network...

That would be very nice, but unfortunately these days with NAT, VPN, etc we are not dealing with just one tcp/ip network :).


Because of the many questions regarding how the network mapping etc. is done i would like to describe another scenario which shows the same problem, but is probably better reproducible:

Thanks, this is helpful.


Use virtualbox on your host machine and install a virtual machine (e.g. running linux). let's name it server1. deploy a service with the generic provider on server1. the service will bind to the local network interface and use the local name of the linux machine: server1. that means e.g. the endpoint ID is ecftcp://server1.local:3282/server .

in virtualbox on the host machine configure a port mapping from port 3282 into the virtual machine with the same port 3282.

from a network perspective you are now able to reach the service in the linux box from your host machine using tcp localhost:3282.

if you now start a service consumer on your host machine which uses the endpoint ecftcp://localhost:3282/server or the real hostname of the host machine e.g. ecftcp://scotts-machine:3282/server you will get a succesful tcp/ip connection between client and server, but (of course) the service import is not working because of the different endpoint id's ....

Right.ÂÂ See explanation above for why this is the case with a group model/generic distribution provider.

So, if you can give up the group model and the ability to export services from peers...and it sounds like you can...then you should be able to give one of the JaxRSProviders a try.

If tcp is needed for your required performance, then to be safe, I would suggest trying a very very small Java application to create a socket connection, read and write a few bytes and make sure that your target network will allow such client-server comm, as many VPN networks limit traffic by port and/or protocol, etc. This is one reason it's so hard for me or anyone to create a 'general' socket provider that will work on any VPN, NAT, network, etc. They can be configured in ways that will allow some things (e.g. odbc, http over specific ports) and not allow others (e.g. socket comm over port xxxx).

If the Java socket app works in your target environment then I would suggest creating a very simple new distribution provider starting with the trivial provider [3].ÂÂ If you decide to do this I can/will help and would welcome it, but for full-effort on it I would need some additional arrangement.


i need a solution for that without adding local hostname entries, dns changes and additional servers or infrastructure ;)

Ok.ÂÂÂ

Regards,

Scott

[1] https://github.com/ECF/kryo-serialization
[2] https://github.com/ECF
[3] http://git.eclipse.org/c/ecf/org.eclipse.ecf.git/tree/examples/bundles/org.eclipse.ecf.examples.provider.trivial