zookeeper-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Rakesh R <rake...@huawei.com>
Subject RE: adding a separate thread to detect network timeouts faster
Date Thu, 12 Sep 2013 07:05:13 GMT

AFAIK, ping requests would not involve any disk I/O, but it would go through the RequestProcessor
chain and executes sequentially. 
There could be cases when there are another set of requests which are in the queue for committing(say
these requests needs database/disk operations). Now a ping request has come from the client,
this will be queued up at the end of the queue. In this case, it would delay the ping request
processing and resulting in slow responses. 

Here the server is slow due to I/O response time and affecting the client ping responses.
Anyway after seeing the ping failure, client would look for another server.

Earlier I tried by passing ping requests from entering to RequestProcessor chain, instead
directly send response back to the client. It has disadvantage of violating the requests lifecycle.
Interesting point is how to differentiate the slow servers and servers which are really down...


-----Original Message-----
From: mutsuzaki@gmail.com [mailto:mutsuzaki@gmail.com] On Behalf Of Michi Mutsuzaki
Sent: 12 September 2013 02:07
To: user@zookeeper.apache.org
Cc: German Blanco
Subject: Re: adding a separate thread to detect network timeouts faster

Slow disk does affect client <-> server ping requests since ping requests go through
the commit processor.

Here is how the current client <-> server ping request works. Say the session timeout
is set to 30 seconds.

1. The client sends a ping request if the session has been inactive for 10 seconds (1/3 of
the session timeout).
2. The client waits for ping response for another 10 seconds (1/3 of the session timeout).
3. If the client doesn't receive ping response after 10 seconds, it tries to connect to another

So in this case, it can take up to 20 seconds for the client to detect a server failure. I
think this 1/3 value is picked somewhat arbitrarily. Maybe you can make this configurable
for faster failure detection instead of introducing another heartbeat mechanism?


On Tue, Sep 10, 2013 at 11:32 PM, Jeremy Stribling <strib@nicira.com> wrote:
> Hi Germán,
> A very quick scan of that JIRA makes me think you're talking about
> server->server heartbeats, and not client->server heartbeats (which is 
> server->what
> I'm talking about).  I have not tested it explicitly or inspected that 
> part of the code, but I've hit many cases in testing and production 
> where client session expirations coincide with long fsync times as logged by the server.
> Jeremy
> On 09/10/2013 10:40 PM, German Blanco wrote:
>> Hello Jeremy and all,
>> my idea was that the current implementation of ping handling already 
>> does not wait on disk IO.
>> I am even working in a JIRA case that is related with this:
>> https://issues.apache.org/jira/browse/ZOOKEEPER-87
>> And I have also made some tests that seem to confirm that ping 
>> handling is done in a different thread than transaction handling.
>> But actually, I don't have any confirmation from any person in this 
>> project. Are you sure that ping handling waits on IO for anything? 
>> Have you tested it?
>> Regards,
>> Germán Blanco.
>> On Tue, Sep 10, 2013 at 11:05 PM, Jeremy Stribling <strib@nicira.com>
>> wrote:
>>> Good suggestion, thanks.  At the very least, I think what we have in 
>>> mind would be off by default, so users could only turn it on if they 
>>> know they have relatively few clients and slow disks.  An adaptive 
>>> scheme would be even better, obviously.
>>> On 09/10/2013 02:04 PM, Ted Dunning wrote:
>>>> Perhaps you should be suggesting a design that is adaptive rather 
>>>> than configured and guarantees low overhead at the cost of 
>>>> notification time in extreme scenarios.
>>>> For instance, the server can send no more than 1000 (or whatever 
>>>> number) HB's per second and never more than one per second to any 
>>>> client.  This caps the cost nicely.
>>>> On Tue, Sep 10, 2013 at 1:59 PM, Ted Dunning
>>>> <ted.dunning@gmail.com<mailto:
>>>> ted.dunning@gmail.com>**> wrote:
>>>>      Since you are talking about client connection failure detection,
>>>>      no, I don't think that there is a major barrier other than
>>>>      actually implementing a reliable check.
>>>>      Keep in mind the cost.  There are ZK installs with 100,000
>>>>      clients.  If these are heartbeating every 2 seconds, you have
>>>>      50,000 packets per second hitting the quorum or 10,000 per server
>>>>      if all connections are well balanced.
>>>>      If you only have 10 clients, the network burden is nominal.
>>>>      On Tue, Sep 10, 2013 at 1:34 PM, Jeremy Stribling
>>>>      <strib@nicira.com <mailto:strib@nicira.com>> wrote:
>>>>          I mostly agree, but let's assume that a ~5x speedup in
>>>>          detecting those types of failures is considered significant
>>>>          for some people. Are there technical reasons that would
>>>>          prevent this idea from working?
>>>>          On 09/10/2013 01:31 PM, Ted Dunning wrote:
>>>>              I don't see the strong value here.  A few failures would
>>>>              be detected more
>>>>              quickly, but I am not convinced that this would actually
>>>>              improve
>>>>              functionality significantly.
>>>>              On Tue, Sep 10, 2013 at 1:01 PM, Jeremy Stribling
>>>>              <strib@nicira.com <mailto:strib@nicira.com>> wrote:
>>>>                  Hi all,
>>>>                  Let's assume that you wanted to deploy ZK in a
>>>>                  virtualized environment,
>>>>                  despite all of the known drawbacks.  Assume we could
>>>>                  deploy it such that
>>>>                  the ZK servers were all using independent CPUs and
>>>>                  storage (though not
>>>>                  dedicated disks).  Obviously, the shared disks (shared
>>>>                  with other, non-ZK
>>>>                  VMs on the same hypervisor) will cause ZK to hit the
>>>>                  default session
>>>>                  timeout occasionally, so you would need to raise the
>>>>                  existing session
>>>>                  timeout to something like 30 seconds.
>>>>                  I'm curious if there would be any technical drawbacks
>>>>                  to adding an
>>>>                  additional heartbeat mechanism between the clients and
>>>>                  the servers, which
>>>>                  would have the goal of detecting network-only failures
>>>>                  faster than the
>>>>                  existing heartbeat mechanism.  The idea is that there
>>>>                  would be a new thread
>>>>                  dedicated to processing these heartbeats, which would
>>>>                  not get blocked on
>>>>                  I/O.  Then the clients could configure a second,
>>>>                  smaller timeout value, and
>>>>                  it would be assumed that any such timeout indicated a
>>>>                  real problem.  The
>>>>                  existing mechanism would still be in place to catch
>>>>                  I/O-related errors.
>>>>                  I understand the philosophy that there should be some
>>>>                  heartbeat mechanism
>>>>                  that takes the disk into account, but I'm having
>>>>                  trouble coming up with
>>>>                  technical reasons not to add a second mechanism.
>>>>                  Obviously, the advantage
>>>>                  would be that the clients could detect network
>>>>                  failures and system crashes
>>>>                  more quickly in an environment with slow disks, and
>>>>                  fail over to other
>>>>                  servers more quickly.  The only disadvantages I can
>>>>                  come up with are:
>>>>                  1) More code complexity, and slightly more heartbeat
>>>>                  traffic on the wire
>>>>                  2) I think the servers have to log session expirations
>>>>                  to disk, so if the
>>>>                  sessions expire at a faster rate than the disk can
>>>>                  handle, it might lead to
>>>>                  a large backlog.
>>>>                  Are there other drawbacks I am missing?  Would a patch
>>>>                  that added
>>>>                  something like this be considered, or is it dead from
>>>>                  the start? Thanks,
>>>>                  Jeremy

View raw message