spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Matei Zaharia <matei.zaha...@gmail.com>
Subject Re: executor failures w/ scala 2.10
Date Fri, 01 Nov 2013 21:59:22 GMT
Hey Imran,

Good to know that Akka 2.1 handles this — that at least will give us a start.

In the old code, executors certainly did get flagged as “down” occasionally, but that
was due to a timeout we controlled (we keep sending heartbeats back and forth to track them).
The timeout used to be smaller and usually the reason to exceed it was GC. However, if Akka
2.2 can sometimes drop the connections itself, this is a problem and we either have to use
the reliable proxies for everything or see if we can configure it otherwise. Anyway, we’ll
definitely look into it.

Matei

On Nov 1, 2013, at 1:09 PM, Imran Rashid <imran@quantifind.com> wrote:

> I downgraded spark to akka 2.1.0, and everything seems to work now.  I'm going to run
my tests a few more times , but I'd really have expected to see a failure by now w/ the 2.2.3
version.
> 
> I'll submit a patch shortly (need to fix some compile errors in streaming still).
> 
> Matei -- I think I realize now that when you were talking about the expectation of a
tcp connection staying alive, you were explaining why this is *not* a bug in the current release.
 You wouldn't end up in a situation where the executor thinks it finished the task, but the
driver doesn't know about it, b/c if the connection dies, the executor wil get restarted.
 That makes sense.  But, it seems like if we upgrade to akka 2.2.x, a lot of things change.
 I was probably wrong about seeing that problem in previous releases -- it was just a vague
recollection, which fit my current theories, so I jumped to conclusions.
> 
> thanks everyone
> 
> 
> 
> On Fri, Nov 1, 2013 at 9:27 AM, Imran Rashid <imran@quantifind.com> wrote:
> thanks everyone for all of the input.
> 
> Matei: makes a lot more sense with your explanation of spark's expected behavior of tcp,
I can see why this makes sense now.  But, to show my total ignorance here, I'm wondering that
when the connection does break, are you sure all of your messages that you thought you sent
before the break were received?  I'm guessing that you don't.  Which is fine, if the response
to that is to have the executor just die completely, and restart.  that was the behavior I
was initially observing with the code on the 2.10 branch, where the executor handles a DisassociatedEvent
explicitly, and dies.
> 
> But -- is that the behavior we want?  do we want it to be robust to tcp connections breaking,
without having to completely restart the executor?  you might say that dying & restarting
will lead to correct behavior, even if its inefficient.  But sometimes, I've seen restarts
so frequently that no progress is made.
> 
> I don't see why this changed w/ the different versions of akka -- I don't see any relevant
configuration settings that would change how "strongly" tcp tries to keep the connection alive,
but I may be missing something.  But it does seem like the netty configuration options have
changed completely between the two versions:
> http://doc.akka.io/docs/akka/2.2.3/scala/remoting.html#Remote_Configuration
> vs
> http://doc.akka.io/docs/akka/2.0.5/scala/remoting.html
> 
> btw, akka 2.1.0 also has been built for scala 2.10:
> http://search.maven.org/#artifactdetails|com.typesafe.akka|akka-remote_2.10|2.1.0|bundle
> and its netty configuration is closer to 2.0.5:
> http://doc.akka.io/docs/akka/2.1.0/scala/remoting.html
> 
> perhaps someone more knowledge then me about netty & tcp can look through the changes
and decide what the right changes are.
> 
> Prashant said:
> >Before we conclude something about reliable messaging, I want you to for once consider
other possibilities like >actual network reconnection and may be a GC pause ? Try connecting
something like jconsole (or alike ) and >see what happens on the driver and executor.
> >
> >My doubt are since we are using standalone mode where even master and worker are
also actors then if we see >a weird behaviour on the executor and driver then Why not on
master and worker too ? They should also break >away from each other. For this reason,
I am doubting our conclusions and may be if we narrow down the >problem first before we
conclude something. It is a regression in akka 2.2.3 it uses more memory than it used to >be
in 2.1.x.  
> >See https://github.com/akka/akka/issues/1810 
> 
> 
> Well, there could easily be the same problem with dropped connections between master
& worker -- they just communicate so little, it doesn't really matter.  The odds that
a message gets dropped between them is very low, only because there are barely any messages.
> 
> I completely agree that the problem could be because of a contention, or gc pause, etc.
 In fact, I'm only giving spark 24 out of 32 cores available on each box, and 90g out of 125g
memory.  I've looked at gc a little with jstat, and I did see some gc pauses but nothing ridiculous.
> 
> But, I think the question remains.  Suppose it is gc pauses, etc. that cause the disassociation
events; what do we do to fix it?  How can we diagnose the problem, and figure out which of
the configuration variables to tune?  clearly, there *will be* long gc pauses, and the networking
layer needs to be able to deal with them.
> 
> still I understand your desire to see if that might be the cause of the problem in this
particular case, so I will dig a little more.
> 
> 
> (btw, should I move this thread to the dev list now?  it is getting into the nitty-gritty
of implementation ...)
> 
> On Fri, Nov 1, 2013 at 1:15 AM, Matei Zaharia <matei.zaharia@gmail.com> wrote:
> Yes, so far they’ve been built on that assumption — not that Akka would *guarantee*
delivery in that as soon as the send() call returns you know it’s delivered, but that Akka
would act the same way as a TCP socket, allowing you to send a stream of messages in order
and hear when the connection breaks. Maybe that isn’t what they want to provide, but I'd
find it weird, because it’s very easy to write a server with this property.
> 
> Matei
> 
> On Oct 31, 2013, at 9:58 PM, Sriram Ramachandrasekaran <sri.rams85@gmail.com> wrote:
> 
>> Sorry if I my understanding is wrong. May be, for this particular case it might be
something to do with the load/network, but, in general, are you saying that, we build these
communication channels(block manager communication, task events communication, etc) assuming
akka would take care of it? I somehow feel that, it's being overly optimistic. Correct me
if I am wrong.
>> 
>> 
>> 
>> On Fri, Nov 1, 2013 at 10:08 AM, Matei Zaharia <matei.zaharia@gmail.com> wrote:
>> It’s true that Akka’s delivery guarantees are in general at-most-once, but if
you look at the text there it says that they differ by transport. In the previous version,
I’m quite sure that except maybe in very rare circumstances or cases where we had a bug,
Akka’s remote layer always kept connections up between each pair of hosts. So the guarantee
was that as long as you haven’t received a “disconnected” event, your messages are being
delivered, though of course when you do receive that event you don’t know which messages
have really made it through unless you acked them. But that didn’t matter for our use case
— from our point of view an executor was either up or down.
>> 
>> For this reason I still think it should be possible to configure Akka to do the same
on 2.2. Most likely some timeouts just got lower. With large heaps you can easily get a GC
pause of 60 seconds, so these timeouts should be in the minutes.
>> 
>> If for some reason this isn’t the case, then we have a bigger problem — there
are *lots* of messages beyond task-finished that need to be sent reliably, including things
like block manager events (a block was added / removed on this node) and commands to tell
the block manager to drop data. It would be silly to implement acks at the application level
for all these. But I doubt this is the case. Prashant’s observation that the standalone
cluster manager stayed up is a further sign that this might be due to GC.
>> 
>> Matei
>> 
>> On Oct 31, 2013, at 9:11 PM, Sriram Ramachandrasekaran <sri.rams85@gmail.com>
wrote:
>> 
>>> Hi Imran,
>>> Just to add, we've noticed dis-associations in a couple projects that we built(using
akka 2.2.x not spark). We went to some details to find out what was happening. As Matei, suggested,
Akka keeps the TCP connection open and uses that to talk to peers. We noticed that in our
case, initially, we were seeing dis-associations generally at the end of keep-alive duration.
So, when the keep-alive duration ends, at the TCP layer, a keep-alive probe gets sent to inform
the peer on the other side that the connection is still alive/valid. For some reason, the
probe dint renew the keep-alive connection and we saw a lot of dis-associations during that
time. Later, we realized this was not a pattern either. This thread contains the full history
of our discussions with the Akka team. It's still open and unclear as to what was causing
it for our case. 
>>> We tried tweaking various settings of akka(wrt heartbeats, failure detector,
even plugged-in our own failure detector with no effect).
>>> 
>>> Imran - Just to clarify your point on message delivery - akka's message delivery
policy is at-most-once. However, there's no guarantee for a message to be delivered to a peer.
The documentation clearly explains that. http://doc.akka.io/docs/akka/2.0.2/general/message-send-semantics.html.
It's the responsibility of the application developer to handle cases where message is suspected
to be not have been delivered. 
>>> 
>>> I hope this helps.
>>> 
>>> 
>>> 
>>> 
>>> On Fri, Nov 1, 2013 at 8:35 AM, Imran Rashid <imran@quantifind.com> wrote:
>>> 
>>> unfortunately that change wasn't the silver bullet I was hoping for.  Even with
>>> 1) ignoring DisassociatedEvent
>>> 2) executor uses ReliableProxy to send messages back to driver
>>> 3) turn up akka.remote.watch-failure-detector.threshold=12
>>> 
>>> 
>>> there is a lot of weird behavior.  First, there are a few DisassociatedEvents,
but some that are followed by AssociatedEvents, so that seems ok.  But sometimes the re-associations
are immediately followed by this:
>>> 
>>> 13/10/31 18:51:10 INFO executor.StandaloneExecutorBackend: got lifecycleevent:
AssociationError [akka.tcp://sparkExecutor@<executor>:41441] -> [akka.tcp://spark@<driver>:41321]:
Error [Invalid address: akka.tcp://spark@<driver>:41321] [
>>> akka.remote.InvalidAssociation: Invalid address: akka.tcp://spark@<driver>:41321
>>> Caused by: akka.remote.transport.Transport$InvalidAssociationException: The remote
system has quarantined this system. No further associations to the remote system are possible
until this system is restarted.
>>> ]
>>> 
>>> On the driver, there are messages like:
>>> 
>>> [INFO] [10/31/2013 18:51:07.838] [spark-akka.actor.default-dispatcher-3] [Remoting]
Address [akka.tcp://sparkExecutor@<executor>:46123] is now quarantined, all messages
to this address will be delivered to dead letters.
>>> [WARN] [10/31/2013 18:51:10.845] [spark-akka.actor.default-dispatcher-20] [akka://spark/system/remote-watcher]
Detected unreachable: [akka.tcp://sparkExecutor@<executor>:41441]
>>> 
>>> 
>>> and when the driver does decide that the executor has been terminated, it removes
the executor, but doesn't start another one.
>>> 
>>> there are a ton of messages also about messages to the block manager master ...
I'm wondering if there are other parts of the system that need to use a reliable proxy (or
some sort of acknowledgement).
>>> 
>>> I really don't think this was working properly even w/ previous versions of spark
/ akka.  I'm still learning about akka, but I think you always need an ack to be confident
w/ remote communicate.  Perhaps the old version of akka just had more robust defaults or something,
but I bet it could still have the same problems.  Even before, I have seen the driver thinking
there were running tasks, but nothing happening on any executor -- it was just rare enough
(and hard to reproduce) that I never bothered looking into it more.
>>> 
>>> I will keep digging ...
>>> 
>>> On Thu, Oct 31, 2013 at 4:36 PM, Matei Zaharia <matei.zaharia@gmail.com>
wrote:
>>> BTW the problem might be the Akka failure detector settings that seem new in
2.2: http://doc.akka.io/docs/akka/2.2.3/scala/remoting.html
>>> 
>>> Their timeouts seem pretty aggressive by default — around 10 seconds. This
can easily be too little if you have large garbage collections. We should make sure they are
higher than our own node failure detection timeouts.
>>> 
>>> Matei
>>> 
>>> On Oct 31, 2013, at 1:33 PM, Imran Rashid <imran@quantifind.com> wrote:
>>> 
>>>> pretty sure I found the problem -- two problems actually.  And I think one
of them has been a general lurking problem w/ spark for a while.
>>>> 
>>>> 1)  we should ignore disassociation events, as you suggested earlier.  They
seem to just indicate a temporary problem, and can generally be ignored.  I've found that
they're regularly followed by AssociatedEvents, and it seems communication really works fine
at that point.
>>>> 
>>>> 2) Task finished messages get lost.  When this message gets sent, we dont'
know it actually gets there:
>>>> 
>>>> https://github.com/apache/incubator-spark/blob/scala-2.10/core/src/main/scala/org/apache/spark/executor/StandaloneExecutorBackend.scala#L90
>>>> 
>>>> (this is so incredible, I feel I must be overlooking something -- but there
is no ack somewhere else that I'm overlooking, is there??)  So, after the patch, spark wasn't
hanging b/c of the unhandled DisassociatedEvent.  It hangs b/c the executor has sent some
taskFinished messages that never get received by the driver.  So the driver is waiting for
some tasks to finish, but the executors think they are all done.
>>>> 
>>>> I'm gonna add the reliable proxy pattern for this particular interaction
and see if its fixes the problem
>>>> http://doc.akka.io/docs/akka/2.2.3/contrib/reliable-proxy.html#introducing-the-reliable-proxy
>>>> 
>>>> imran
>>> 
>>> 
>>> 
>>> -- 
>>> It's just about how deep your longing is!
>> 
>> 
>> 
>> 
>> -- 
>> It's just about how deep your longing is!
> 


Mime
View raw message