Well, this did work when I tried it on clients connected to follower nodes in the cluster. The followers still propagate back their touch table to the leader when they ping, correct? And then the leader does the same logic as when it touches a session? As long as the SessionTrackerImpl lines 173-176 are commented out (in the 3.3 branch), you can touch with a lowered expire time, and so long as that session does not re-connect during that shortened interval, the leader will properly time out the session. Again, I haven't thought through this logic extremely closely so it is possible I am missing some implications (like, what happens when two clients are connected to the same session on different servers, but then, what in the world does that even mean?), but it does work in the basic case so far as my testing has shown. I'll try to get the go-ahead to submit a ticket with the code so you can look at it yourself.
I agree that if a network glitch causes a bunch of live clients to appear to close their sockets on the server they are connected to, you could have problems. This could be somewhat remediated by setting the minsessiontimeout to a value that allows for some slop, but not the amount of slop you want to allow for something like a full gc on a client. But what I'm not sure about is the glitch that actually causes this sort of observed behavior. I am not a networking expert, but in my experience I've seen network glitches that cause sockets to appear to be live that are actually dead, but not vice-versa. Can you tell me what would cause a socket closure with otherwise alive client and server? Thanks for the feedback, Camille -----Original Message----- From: Benjamin Reed [mailto:br...@yahoo-inc.com] Sent: Friday, September 10, 2010 4:02 PM To: Fournier, Camille F. [Tech] Cc: 'zookeeper-user@hadoop.apache.org' Subject: Re: closing session on socket close vs waiting for timeout the problem is that followers don't track session timeouts. they track when they last heard from the sessions that are connected to them and they periodically propagate this information to the leader. the leader is the one that expires the session. your technique only works when the client is connected to the leader. one thing you can do is generate a close request for the socket and push that through the system. that will cause it to get propagated through the followers and processed at the leader. it would also allow you to get your functionality without touching the processing pipeline. the thing that worries me about this functionality in general is that network anomalies can cause a whole raft of sessions to get expired in this way. for example, you have 3 servers with load spread well; there is a networking glitch that cause clients to abandon a server; suddenly 1/3 of your clients will get expired sessions. ben On 09/10/2010 12:17 PM, Fournier, Camille F. [Tech] wrote: > Ben, could you explain a bit more why you think this won't work? I'm trying > to decide if I should put in the work to take the POC I wrote and complete > it, but I don't really want to waste my time if there's a fundamental reason > it's a bad idea. > > Thanks, > Camille > > -----Original Message----- > From: Benjamin Reed [mailto:br...@yahoo-inc.com] > Sent: Wednesday, September 08, 2010 4:03 PM > To: zookeeper-user@hadoop.apache.org > Subject: Re: closing session on socket close vs waiting for timeout > > unfortunately, that only works on the standalone server. > > ben > > On 09/08/2010 12:52 PM, Fournier, Camille F. [Tech] wrote: >> This would be the ideal solution to this problem I think. >> Poking around the (3.3) code to figure out how hard it would be to >> implement, I figure one way to do it would be to modify the session timeout >> to the min session timeout and touch the connection before calling close >> when you get certain exceptions in NIOServerCnxn.doIO. I did this (removing >> the code in touch session that returns if the tickTime is greater than the >> expire time) and it worked (in the standalone server anyway). Interesting >> solution, or total hack that will not work beyond most basic test case? >> >> C >> >> (forgive lack of actual code in this email) >> >> -----Original Message----- >> From: Ted Dunning [mailto:ted.dunn...@gmail.com] >> Sent: Tuesday, September 07, 2010 1:11 PM >> To: zookeeper-user@hadoop.apache.org >> Cc: Benjamin Reed >> Subject: Re: closing session on socket close vs waiting for timeout >> >> This really is, just as Ben says a problem of false positives and false >> negatives in detecting session >> expiration. >> >> On the other hand, the current algorithm isn't really using all the >> information available. The current algorithm is >> using time since last client initiated heartbeat. The new proposal is >> somewhat worse in that it proposes to use >> just the boolean "has-TCP-disconnect-happened". >> >> Perhaps it would be better to use multiple features in order to decrease >> both false positives and false negatives. >> >> For instance, I could imagine that we use the following features: >> >> - time since last client hearbeat or disconnect or reconnect >> >> - what was the last event? (a heartbeat or a disconnect or a reconnect) >> >> Then the expiration algorithm could use a relatively long time since last >> heartbeat and a relatively short time since last disconnect to mark a >> session as disconnected. >> >> Wouldn't this avoid expiration during GC and cluster partition and cause >> expiration quickly after a client disconnect? >> >> >> On Mon, Sep 6, 2010 at 11:26 PM, Patrick Hunt<ph...@apache.org> wrote: >> >> >>> That's a good point, however with suitable documentation, warnings and such >>> it seems like a reasonable feature to provide for those users who require >>> it. Used in moderation it seems fine to me. Perhaps we also make it >>> configurable at the server level for those administrators/ops who don't >>> want >>> to deal with it (disable the feature entirely, or only enable on particular >>> servers, etc...). >>> >>> Patrick >>> >>> On Mon, Sep 6, 2010 at 2:10 PM, Benjamin Reed<br...@yahoo-inc.com> wrote: >>> >>> >>>> if this mechanism were used very often, we would get a huge number of >>>> session expirations when a server fails. you are trading fast error >>>> detection for the ability to tolerate temporary network and server >>>> >>> outages. >>> >>>> to be honest this seems like something that in theory sounds like it will >>>> work in practice, but once deployed we start getting session expirations >>>> >>> for >>> >>>> cases that we really do not want or expect. >>>> >>>> ben >>>> >>>> >>>> On 09/01/2010 12:47 PM, Patrick Hunt wrote: >>>> >>>> >>>>> Ben, in this case the session would be tied directly to the connection, >>>>> we'd explicitly deny session re-establishment for this session type (so >>>>> 4 would fail). Would that address your concern, others? >>>>> >>>>> Patrick >>>>> >>>>> On 09/01/2010 10:03 AM, Benjamin Reed wrote: >>>>> >>>>> >>>>> >>>>>> i'm a bit skeptical that this is going to work out properly. a server >>>>>> may receive a socket reset even though the client is still alive: >>>>>> >>>>>> 1) client sends a request to a server >>>>>> 2) client is partitioned from the server >>>>>> 3) server starts trying to send response >>>>>> 4) client reconnects to a different server >>>>>> 5) partition heals >>>>>> 6) server gets a reset from client >>>>>> >>>>>> at step 6 i don't think you want to delete the ephemeral nodes. >>>>>> >>>>>> ben >>>>>> >>>>>> On 08/31/2010 01:41 PM, Fournier, Camille F. [Tech] wrote: >>>>>> >>>>>> >>>>>> >>>>>>> Yes that's right. Which network issues can cause the socket to close >>>>>>> without the initiating process closing the socket? In my limited >>>>>>> experience in this area network issues were more prone to leave dead >>>>>>> sockets open rather than vice versa so I don't know what to look out >>>>>>> for. >>>>>>> >>>>>>> Thanks, >>>>>>> Camille >>>>>>> >>>>>>> -----Original Message----- >>>>>>> From: Dave Wright [mailto:wrig...@gmail.com] >>>>>>> Sent: Tuesday, August 31, 2010 1:14 PM >>>>>>> To: zookeeper-user@hadoop.apache.org >>>>>>> Subject: Re: closing session on socket close vs waiting for timeout >>>>>>> >>>>>>> I think he's saying that if the socket closes because of a crash (i.e. >>>>>>> not a >>>>>>> normal zookeeper close request) then the session stays alive until the >>>>>>> session timeout, which is of course true since ZK allows reconnection >>>>>>> and >>>>>>> resumption of the session in case of disconnect due to network issues. >>>>>>> >>>>>>> -Dave Wright >>>>>>> >>>>>>> On Tue, Aug 31, 2010 at 1:03 PM, Ted Dunning<ted.dunn...@gmail.com> >>>>>>> wrote: >>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>>> That doesn't sound right to me. >>>>>>>> >>>>>>>> Is there a Zookeeper expert in the house? >>>>>>>> >>>>>>>> On Tue, Aug 31, 2010 at 8:58 AM, Fournier, Camille F. [Tech]< >>>>>>>> camille.fourn...@gs.com> wrote: >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>>> I foolishly did not investigate the ZK code closely enough and it >>>>>>>>> seems >>>>>>>>> that closing the socket still waits for the session timeout to >>>>>>>>> remove the >>>>>>>>> session. >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>> >>>>>> >>>>> >>>> >>>
