From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <vdavydov.dev@gmail.com>
Date: Wed, 17 Oct 2018 11:35:01 +0300
From: Vladimir Davydov <vdavydov.dev@gmail.com>
Subject: Re: [tarantool-patches] Re: [PATCH v2] iproto: introduce a proxy
 module.
Message-ID: <20181017083500.kznvogmbghukee74@esperanza>
References: <20181002180554.1142-1-sergepetrenko@tarantool.org>
 <20181008164807.ewiv7lwfh4e4decg@esperanza>
 <20181016183958.GE5454@chai>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20181016183958.GE5454@chai>
To: Konstantin Osipov <kostja@tarantool.org>
Cc: tarantool-patches@freelists.org, Serge Petrenko <sergepetrenko@tarantool.org>
List-ID: <tarantool-patches.dev.tarantool.org>

On Tue, Oct 16, 2018 at 09:39:58PM +0300, Konstantin Osipov wrote:
> * Vladimir Davydov <vdavydov.dev@gmail.com> [18/10/11 09:55]:
> > On Tue, Oct 02, 2018 at 09:05:54PM +0300, Serge Petrenko wrote:
> > > Since salt proxy sends to a client differs from the salt it recieves
> > > from a remote instance, forwarding auth requests to establish non-guest
> > > connections is a little bit tricky:
> > > let hash1 = sha1(password),
> > >     hash2 = sha1(hash1)
> > > then upon auth proxy recieves such a string from the client:
> > >     reply = xor(hash1, sha1(proxy_salt, hash2))
> > > proxy has to send an auth request of such form to an instance:
> > >     request = xor(hash1, sha1(instance_salt, hash2))
> > > proxy fetches hash2 via a special message to tx thread (again, it is
> > > accessible, since proxy is run on one of the cluster instances).
> > > Then proxy computes hash1 = xor(reply, sha1(proxy_salt, hash2)) and
> > > computes the request using hash1, hash2 and instance_salt.
> > 
> > So unless the user is fine with guest access (which is rather unlikely
> > AFAIU), it doesn't make sense to run a proxy on a standalone instance,
> > does it?
> > 
> > If so, may be we could simplify both configuration and the code by
> > requiring a proxy to be a part of the replica set?
> 
> It's OK to simplify the patch now and only make it work within a
> replica set. In future we will have replication groups, remember,
> and this way we'll get a standalone proxy for free.
> 
> > I mean instead of netbox.listen(), we could add a knob to box.cfg, say
> > box.cfg.proxy_enable = true|false. If this knob was set, the instance
> > would automatically forward all incoming iproto requests to members of
> > the replica set (including self). What do you think?
> 
> We need a new knob for two reasons:
> - we need to be able to turn on listen port before box.cfg; it's
>   albeit separate, a popular user request.

Then it should be done in a separate patch before introducing proxy.

> - we need backward compatibility. i.e. old listen should work the
>   old way, to not surprise users

How? Two iproto threads?

> > 
> > > Proxy may be configured like this:
> > > ```
> > > netbox = require("net.box")
> > > netbox.listen(uri_to_listen, {cluster={
> > > 	{uri=uri1, is_master=false},
> > > 	{uri=uri2, is_master=true},
> > > 	...
> > > 	}})
> > > ```
> > 
> > I don't like that the user has to explicitly configure which participant
> > is rw and which is ro. How will it work when box.ctl.promote is finally
> > implemented?
> 
> Can we please implement a proxy in which there is no
> {cluster=} option? As we were discussing with Vlad, proxy should
> be integrated with box.ctl.promote() and learn about ro/rw masters
> from it.

Agree, but how can we implement proxy before promote then? May be, we
could use IPROTO_VOTE to inquire which replica is rw and which is ro?