----- Original Message -----
> From: "Adam Litke" <a...@us.ibm.com>
> To: "Saggi Mizrahi" <smizr...@redhat.com>
> Cc: "Livnat Peer" <lp...@redhat.com>, engine-de...@ovirt.org, 
> vdsm-devel@lists.fedorahosted.org
> Sent: Thursday, January 26, 2012 1:58:40 PM
> Subject: Re: [vdsm] [Engine-devel] [RFC] New Connection Management API
> 
> On Thu, Jan 26, 2012 at 10:00:57AM -0500, Saggi Mizrahi wrote:
> > <snip>
> > Again trying to sum up and address all comments
> > 
> > Clear all:
> > ==========
> > My opinions is still to not implement it.
> > Even though it might generate a bit more traffic premature
> > optimization is bad and there are other reasons we can improve
> > VDSM command overhead without doing this.
> > 
> > In any case this argument is redundant because my intention is (as
> > Litke pointed out) is to have a lean API.
> > and API call is something you have to support across versions, this
> > call implemented in the engine is something that no one has to
> > support and can change\evolve easily.
> > 
> > As a rule, if an API call C and be implemented by doing A + B then
> > C is redundant.
> > 
> > List of connections as args:
> > ============================
> > Sorry I forgot to respond about that. I'm not as strongly opposed
> > to the idea as the other things you suggested. It'll just make
> > implementing the persistence logic in VDSM significantly more
> > complicated as I will have to commit multiple connection
> > information to disk in an all or nothing mode. I can create a
> > small sqlitedb to do that or do some directory tricks and exploit
> > FS rename atomicity but I'd rather not.
> 
> I would be strongly opposed to introducing a sqlite database into
> vdsm just to
> enable "convenience mode" for this API.  Does the operation really
> need to be
> atomic?  Why not just perform each connection sequentially and return
> a list of
> statuses? Is the only motivation for allowing a list of parameters
> to reduce
> the number of API calls between engine and vdsm)?  If so, the same
> argument
> Saggi makes above applies here.

I try and have VDSM expose APIs that are simple to predict. a command can 
either succeed or fail.
The problem is not actually validating the connections. The problem is that 
once I concluded that they are all OK I need to persist to disk the information 
that will allow me to reconnect if VDSM happens to crash. If I naively save 
them one by one I could get in a state where only some of the connections 
persisted before the operation failed. So I have to somehow put all this in a 
transaction.

I don't have to use sqlite. I could also put all the persistence information in 
a new dir for every call named <UUID>.tmp. Once I wrote everything down I 
rename the directory to just <UUID> and fsync it. This is guarantied by posix 
to be atomic. For unmanage, I move all the persistence information from the 
directories they sit in to a new dir named <UUID>. Rename it to a <UUDI>.tmp, 
fsync it and then remove it.

This all just looks like more trouble then it's worth to me.

> 
> > The demands are not without base. I would like to keep the code
> > simple under the hood in the price of a few more calls. You would
> > like to make less calls and keep the code simpler on your side.
> > There isn't a real way to settle this.
> > If anyone on the list as pros and cons for either way I'd be happy
> > to hear them.
> > If no compelling arguments arise I will let Ayal call this one.
> > 
> > Transient connections:
> > ======================
> > The problem you are describing as I understand it is that VDSM did
> > not respond and not that the API client did not respond.
> > Again, this can happen for a number of reason, most of which VDSM
> > might not be aware that there is actually a problem (network
> > issues).
> > 
> > This relates to the EOL policy. I agree we have to find a good way
> > to define an automatic EOL for resources. I have made my
> > suggestion. Out of the scope of the API.
> > 
> > In the meantime cleaning stale connections is trivial and I have
> > made it clear a previous email about how to go about it in a
> > simple non intrusive way. Clean hosts on host connect, and on
> > every poll if you find connections that you don't like. This
> > should keep things squeaky clean.
> > 
> > ----- Original Message -----
> > > From: "Livnat Peer" <lp...@redhat.com>
> > > To: "Saggi Mizrahi" <smizr...@redhat.com>
> > > Cc: vdsm-devel@lists.fedorahosted.org, engine-de...@ovirt.org
> > > Sent: Thursday, January 26, 2012 5:22:42 AM
> > > Subject: Re: [Engine-devel] [RFC] New Connection Management API
> > > 
> > > On 25/01/12 23:35, Saggi Mizrahi wrote:
> > > > <SNIP>
> > > > This is mail was getting way too long.
> > > > 
> > > > About the clear all verb.
> > > > No.
> > > > Just loop, find the connections YOU OWN and clean them. Even
> > > > though
> > > > you don't want to support multiple clients to VDSM API doesn't
> > > > mean the engine shouldn't behave like a proper citizen.
> > > > It's the same reason why VDSM tries and not mess system
> > > > resources
> > > > it didn't initiate.
> > > 
> > > 
> > > There is a big difference, VDSM living in hybrid mode with other
> > > workload on the host is a valid use case, having more than one
> > > concurrent manager for VDSM is not.
> > > Generating a disconnect request for each connection does not seem
> > > like
> > > the right API to me, again think on the simple flow of moving
> > > host
> > > from
> > > one data center to another, the engine needs to disconnect tall
> > > storage
> > > domains (each domain can have couple of connections associated
> > > with
> > > it).
> > > 
> > > I am giving example from the engine use cases as it is the main
> > > user
> > > of
> > > VDSM ATM but I am sure it will be relevant to any other user of
> > > VDSM.
> > > 
> > > > 
> > > > ------------------------
> > > > 
> > > > As I see it the only point of conflict is the so called
> > > > non-peristed connections.
> > > > I will call them transient connections from now on.
> > > > 
> > > > There are 2 user cases being discussed
> > > > 1. Wait until a connection is made, if it fails don't retry and
> > > > automatically unmanage.
> > > > 2. If the called of the API forgets or fails to unmanage a
> > > > connection.
> > > > 
> > > 
> > > Actually I was not discussing #2 at all.
> > > 
> > > > Your suggestion as I understand it:
> > > > Transient connections are:
> > > >      - Connection that VDSM will only try to connect to once
> > > >      and
> > > >      will not reconnect to in case of disconnect.
> > > 
> > > yes
> > > 
> > > > 
> > > > My problem with this definition that it does not specify the
> > > > "end
> > > > of life" of the connection.
> > > > Meaning it solves only use case 1.
> > > 
> > > since this is the only use case i had in mind, it is what i was
> > > looking for.
> > > 
> > > > If all is well, and it usually is, VDSM will not invoke a
> > > > disconnect.
> > > > So the caller would have to call unmanage if the connection
> > > > succeeded at the end of the flow.
> > > 
> > > agree.
> > > 
> > > > Now, if you are already calling unmanage if connection
> > > > succeeded
> > > > you can just call it anyway.
> > > 
> > > not exactly, an example I gave earlier on the thread was that
> > > VSDM
> > > hangs
> > > or have other error and the engine can not initiate unmanaged,
> > > instead
> > > let's assume the host is fenced (self-fence or external fence
> > > does
> > > not
> > > matter), in this scenario the engine will not issue unmanage.
> > > 
> > > > 
> > > > instead of doing: (with your suggestion)
> > > > ----------------
> > > > manage
> > > > wait until succeeds or lastError has value
> > > > try:
> > > >   do stuff
> > > > finally:
> > > >   unmanage
> > > > 
> > > > do: (with the canonical flow)
> > > > ---
> > > > manage
> > > > try:
> > > >   wait until succeeds or lastError has value
> > > >   do stuff
> > > > finally:
> > > >   unmanage
> > > > 
> > > > This is simpler to do than having another connection type.
> > > 
> > > You are assuming the engine can communicate with VDSM and there
> > > are
> > > scenarios where it is not feasible.
> > > 
> > > > 
> > > > Now that we got that out of the way lets talk about the 2nd use
> > > > case.
> > > 
> > > Since I did not ask VDSM to clean after the (engine) user and you
> > > don't
> > > want to do it I am not sure we need to discuss this.
> > > 
> > > If you insist we can start the discussion on who should implement
> > > the
> > > cleanup mechanism but I'm afraid I have no strong arguments for
> > > VDSM
> > > to
> > > do it, so I rather not go there ;)
> > > 
> > > 
> > > You dropped from the discussion my request for supporting list of
> > > connections for manage and unmanage verbs.
> > > 
> > > > API client died in the middle of the operation and unmanage was
> > > > never called.
> > > > 
> > > > Your suggested definition means that unless there was a problem
> > > > with the connection VDSM will still have this connection
> > > > active.
> > > > The engine will have to clean it anyway.
> > > > 
> > > > The problem is, VDSM has no way of knowing that a client died,
> > > > forgot or is thinking really hard and will continue on in about
> > > > 2
> > > > minutes.
> > > 
> > > > 
> > > > Connections that live until they die is a hard to define and
> > > > work
> > > > with lifecycle. Solving this problem is theoretically simple.
> > > > 
> > > > Have clients hold some sort of session token and force the
> > > > client
> > > > to update it at a specified interval. You could bind resources
> > > > (like domains, VMs, connections) to that session token so when
> > > > it
> > > > expires VDSM auto cleans the resources.
> > > > 
> > > > This kind of mechanism is out of the scope of this API change.
> > > > Further more I think that this mechanism should sit in the
> > > > engine
> > > > since the session might actually contain resources from
> > > > multiple
> > > > hosts and resources that are not managed by VDSM.
> > > > 
> > > > In GUI flows specifically the user might do actions that don't
> > > > even
> > > > touch the engine and forcing it to refresh the engine token is
> > > > simpler then having it refresh the VDSM token.
> > > > 
> > > > I understand that engine currently has no way of tracking a
> > > > user
> > > > session. This, as I said, is also true in the case of VDSM. We
> > > > can
> > > > start and argue about which project should implement the
> > > > session
> > > > semantics. But as I see it it's not relevant to the connection
> > > > management API.
> > > 
> > > 
> > _______________________________________________
> > vdsm-devel mailing list
> > vdsm-devel@lists.fedorahosted.org
> > https://fedorahosted.org/mailman/listinfo/vdsm-devel
> 
> --
> Adam Litke <a...@us.ibm.com>
> IBM Linux Technology Center
> 
> 
_______________________________________________
vdsm-devel mailing list
vdsm-devel@lists.fedorahosted.org
https://fedorahosted.org/mailman/listinfo/vdsm-devel

Reply via email to