kafka-users mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Eno Thereska <eno.there...@gmail.com>
Subject Re: Reliably implementing global KeyValueStore#get
Date Wed, 07 Jun 2017 21:14:53 GMT
Hi Steven,

You are right in principle. The thing is that what we shipped in Kafka is just the low level
bare bones that in a sense belong to Kafka. A middle layer that keeps track of the data is
absolutely needed, and it should hopefully hide the distributed system challenges from the
end user. Now the question is how should such layer look like. I think in all systems there
are some basic assumptions made about frequency of failures and rebalances just to keep the
number of retries sane. I agree with you that in principle a rebalance could be always happening
though.

Eno

> On 6 Jun 2017, at 23:29, Steven Schlansker <sschlansker@opentable.com> wrote:
> 
> 
>> On Jun 6, 2017, at 2:52 PM, Damian Guy <damian.guy@gmail.com> wrote:
>> 
>> Steven,
>> 
>> In practice, data shouldn't be migrating that often. If it is then you
>> probably have bigger problems.
> 
> Understood and agreed, but when designing distributed systems, it usually
> helps to model for the worst case rather than the "well that should never
> happen" case, lest you find yourself fixing those bugs at 3am instead :)
> 
> I'd like to be able to induce extreme pain at the Kafka layer (change leader
> every 3 seconds and migrate all partitions around randomly) and still have
> my app behave correctly.
> 
>> You should be able to use the metadata api
>> to find the instance the key should be on and then when you check that node
>> you can also check with the metadata api that the key should still be on
>> this host. If streams is rebalancing while you query an exception will be
>> raised and you'll need to retry the request once the rebalance has
>> completed.
> 
> Agreed here as well.  But let's assume I have a very fast replication
> setup (assume it takes zero time, for the sake of argument) -- I'm fairly
> sure there's still a race here as this exception only fires *during a migration*
> not *after a migration that may have invalidated your metadata lookup completes*
> 
>> 
>> HTH,
>> Damian
>> 
>> On Tue, 6 Jun 2017 at 18:11 Steven Schlansker <sschlansker@opentable.com>
>> wrote:
>> 
>>> 
>>>> On Jun 6, 2017, at 6:16 AM, Eno Thereska <eno.thereska@gmail.com> wrote:
>>>> 
>>>> Hi Steven,
>>>> 
>>>> Do you know beforehand if a key exists? If you know that and are getting
>>> null() the code will have to retry by refreshing the metadata and going to
>>> the new instance. If you don’t know beforehand if a key exists or not you
>>> might have to check all instances of a store to make sure.
>>>> 
>>> 
>>> No, I am not presupposing that the key can exist -- this is a user visible
>>> API and will
>>> be prone to "accidents" :)
>>> 
>>> Thanks for the insight.  I worry that even checking all stores is not
>>> truly sufficient,
>>> as querying different all workers at different times in the presence of
>>> migrating data
>>> can still in theory miss it given pessimal execution.
>>> 
>>> I'm sure I've long wandered off into the hypothetical, but I dream of some
>>> day being
>>> cool like Jepsen :)
>>> 
>>>> Eno
>>>> 
>>>> 
>>>>> On Jun 5, 2017, at 10:12 PM, Steven Schlansker <
>>> sschlansker@opentable.com> wrote:
>>>>> 
>>>>> Hi everyone, me again :)
>>>>> 
>>>>> I'm still trying to implement my "remoting" layer that allows
>>>>> my clients to see the partitioned Kafka Streams state
>>>>> regardless of which instance they hit.  Roughly, my lookup is:
>>>>> 
>>>>> Message get(Key key) {
>>>>> RemoteInstance instance = selectPartition(key);
>>>>> return instance.get(key); // http remoting
>>>>> }
>>>>> 
>>>>> RemoteInstance.get(Key key) { // http endpoint
>>>>> return readOnlyKeyValueStore.get(key);
>>>>> }
>>>>> 
>>>>> However, the mapping of partitions to instances may change.
>>>>> If you call KeyValueStore.get(K) where K is on a partition you
>>>>> don't own, it returns null.  This is indistinguishable from a
>>>>> successful get on a key that doesn't exist.
>>>>> 
>>>>> If one instance selects a sibling instance right as the partition is
>>> failing
>>>>> off of that instance, it may get routed there and by the time it gets
>>>>> the request no longer "owns" the partition -- returns a false 'null'.
>>>>> 
>>>>> You can try re-checking after you get a null value, but that's
>>> susceptible
>>>>> to the same race -- it's unlikely but possible that the data migrates
>>> *back*
>>>>> before you do this re-check.
>>>>> 
>>>>> Is there any way to correctly implement this without races?  I'd imagine
>>>>> you need a new primitive like KeyValueStore#get that atomically finds
>>>>> the key or throws an exception if it is not in an owned partition
>>>>> at the time of lookup so you know to recheck the partition and retry.
>>>>> 
>>>>> Thoughts?
>>>>> 
>>>>> Thanks again,
>>>>> Steven
>>>>> 
>>>> 
>>> 
>>> 
> 


Mime
View raw message