Guys,
Thanks for the discussion!
Summary:
1. Q: How KAFKA-1367 (isr is inconsistent in brokers' metadata cache) can
affect implementation?
A: We can fix this issue for the leading broker - ReplicaManager (or
Partition)
component should have accurate isr list, then with leading broker
having correct
info, to do a describe-topic we will need to define leading brokers
for partitions
and ask those for a correct isr list.
Also, we should consider adding lag information to TMR for each
follower for
partition reassignment, as Jun suggested above.
2. Q: What if user adds different alter commands for the same topic in scope
of one batch request?
A: Because of the async nature of AlterTopicRequest it will be very
hard then
to "assemble" the expected (in terms of checking whether request is
complete)
result if we let users do this. Also it will be very confusing. It
was proposed not to
let users do this (probably add new Error for such cases).
3. Q: AlterTopicRequest semantics: now when we merged AlterTopic and
ReassingPartitons in which order AlterTopic fields should be
resolved?
A: This item is not clear. There was a proposal to let user change only
one thing at a time, e.g. specify either new Replicas, or
ReplicaAssignment.
This can be a simple solution, but it's a very strict rule. E.g.
currently with
TopicCommand user can increase nr of partitions and define replica
assignment
for newly added partitions. Taking into account item 2. this will
be even harder
for user to achieve this.
4. Q: Do we need such accurate errors returned from the server:
InvalidArgumentPartitions,
InvalidArgumentReplicas etc.
A: I started implementation to add proposed error codes and now I think
probably
InvalidArgumentError should be sufficient. We can do simple
validations on
the client side (e.g. AdminClient can ensure nr of partitions
argument is positive),
others - which can be covered only on server (probably invalid
topic config,
replica assignment includes dead broker etc) - will be done on
server, and in case
of invalid argument we will return InvalidArgumentError without
specifying the
concrete field.
It'd be great if we could cover these remaining issues, looks like they are
minor,
at least related to specific messages, not the overall protocol. - I think
with that I can
update confluence page and update patch to reflect all discussed items.
This patch
will probably include Wire protocol messages and server-side code to handle
new
requests. AdminClient and cli-tool implementation can be the next step.
Thanks,
Andrii Biletskyi
On Wed, Apr 15, 2015 at 7:26 PM, Jun Rao <j...@confluent.io> wrote:
> Andrii,
>
> 500. I think what you suggested also sounds reasonable. Since ISR is only
> maintained accurately at the leader, TMR can return ISR if the broker is
> the leader of a partition. Otherwise, we can return an empty ISR. For
> partition reassignment, it would be useful to know the lag of each
> follower. Again, the leader knows this info. We can probably include that
> info in TMR as well.
>
> 300. I think it's probably reasonable to restrict AlterTopicRequest to
> change only one thing at a time, i.e., either partitions, replicas, replica
> assignment or config.
>
> Thanks,
>
> Jun
>
> On Mon, Apr 13, 2015 at 10:56 AM, Andrii Biletskyi <
> andrii.bilets...@stealth.ly> wrote:
>
> > Jun,
> >
> > 404. Great, thanks!
> >
> > 500. If I understand correctly KAFKA-1367 says ISR part of TMR may
> > be inconsistent. If so, then I believe all admin commands but
> describeTopic
> > are not affected. Let me emphasize that it's about AdminClient
> operations,
> > not about Wire Protocol requests. What I mean:
> > To verify AdminClient.createTopic we will need (consistent) 'topics' set
> > from TMR (we don't need isr)
> > To verify alterTopic - again, probably 'topics' and 'assigned replicas' +
> > configs
> > To verify deleteTopic - only 'topics'
> > To verify preferredReplica - 'leader', 'assigned replicas'
> > To verify reassignPartitions - 'assigned replicas' ? (I'm not sure about
> > this one)
> > If everything above is correct, then AdminClient.describeTopic is the
> only
> > command under risk. We can actually workaround it - find out the leader
> > broker
> > and ask TMR that leading broker to get up-to-date isr list.
> > Bottom line: looks like 1367 is a separate issue, and is not a blocker
> for
> > this
> > KIP. I'm a bit concerned about adding new requests as a must-have part
> > of this KIP when we don't know what we want to include to those requests.
> >
> > Also, I'd like to write down the new AlterTopicRequest semantics (if we
> > decide
> > to include replicas there and merge it with ReassignPartitionsRequest)
> > 300. AlterTopicRequest => [TopicName Partitions Replicas
> ReplicaAssignment
> > [AddedConfigEntry] [DeletedConfig]]
> > The fields are resolved in this sequence:
> > 1. Either partition or replicas is defined:
> > ---1.1. ReplicaAssignment is not defined - generate automatic replica
> > assignment
> > for newly added partitions or for replicas parameter increased
> > ---1.2. ReplicaAssignment is defined - increase topic partitions if
> > 'partitions' defined,
> > reassign partitions according to ReplicaAssignment
> > 2. Neither partition nor replicas is defined:
> > ---2.1. ReplicaAssignment is defined - it's a reassign replicas request
> > ---2.2. ReplicaAssignment is not defined - just change topic configs
> > 3. Config fields are handled always and independently from
> > partitions+replicas/replicaAssingment
> > A bit sophisticated, but should cover all cases. Another option - we can
> > say you can define either partitions+replicas or replicaAssignment.
> >
> > 300.5. There is also a new question related to AlterTopicRequest - should
> > we
> > allow users multiple alter-topic instructions for one topic in one batch?
> > I think if we go this way, user will expect we optimize and group
> requests
> > for one topic, but it will add a lot of burden, especially taken into
> > account
> > async semantics of the AlterTopicRequest. I'd rather return some error
> > code,
> > or ignore all but first. Thoughts?
> >
> > Thanks,
> > Andrii Biletskyi
> >
> > On Mon, Apr 13, 2015 at 6:34 AM, Jun Rao <j...@confluent.io> wrote:
> >
> > > Andrii,
> > >
> > > 404. Jay and I chatted a bit. We agreed to leave createTopicRequest as
> > > async for now.
> > >
> > > There is another thing.
> > >
> > > 500. Currently, we have this issue (KAFKA-1367) that the ISR in the
> > > metadata cache can be out of sync. The reason is that ISR is really
> > > maintained at the leader. We can potentially add a new
> BrokerMetaRequest,
> > > which will return useful stats specific to a broker. Such stats can
> > include
> > > (1) for each partition whose leader is on this broker, the ISR and the
> > lag
> > > (in messages) for each of the followers, (2) space used per partition,
> > (3)
> > > remaining space per log dir (not sure how easy it is to get this info).
> > If
> > > we have this new request, we can probably remove the ISR part from TMR
> > v1.
> > > Currently, the producer/consumer client don't really care about ISR.
> The
> > > admin client will then issue BrokerMetaRequest to find out ISR and
> other
> > > stats.
> > >
> > > Thanks,
> > >
> > > Jun
> > >
> > >
> > > On Tue, Apr 7, 2015 at 12:10 PM, Andrii Biletskyi <
> > > andrii.bilets...@stealth.ly> wrote:
> > >
> > > > Hi all,
> > > >
> > > > A summary of our discussion:
> > > >
> > > > 201. Q: Cluster updates in backward compatible way.
> > > > A: Add topicConfigs map property and change constructor, this
> > > > shouldn't break Consumer/Producer since TMR is used in NetworkClient,
> > > > not directly by Consumer/Producer.
> > > >
> > > > 300. Q: Can we merge AlterTopic and ReassignPartitions requests?
> > > > A: It looks like in terms of Wire Protocol partition
> > reassignment
> > > > can
> > > > be just an application of AlterTopicRequest. On the AdminClient side
> we
> > > can
> > > > split this into two separate methods, if needed.
> > > >
> > > > Some additional items that were added today:
> > > > 400. Q: Do we need ListTopicsRequest, we can use TMR for this
> purpose.
> > > > A: The answer depends on whether we can leverage ListTopics
> in
> > > > consumer/producer, because the only benefit of the ListTopics is
> > > > performance
> > > > optimization, otherwise it doesn't worth it.
> > > >
> > > > 401. Q: AdminClient.topicExists - do we need it?
> > > > A: AdminClient.listTopics should be sufficient.
> > > >
> > > > 402. Review AdminClient API and use separate objects instead of
> > > collections
> > > > for methods arguments / return results (e.g. preferredReplica accepts
> > > > Map<String, List<Int>>
> > > > might be better to add separate java object)
> > > >
> > > > 403. Error number in KIP-4 (100x). Currently there are no dedicated
> > > ranges
> > > > for errors, we will probably continue doing it this way.
> > > >
> > > > 404. There were some concerns again about the asynchronous semantics
> > > > of the admin requests. Jun and Jay to agree separately how we want
> > > > to handle it.
> > > >
> > > > Please add / correct me if I missed something.
> > > >
> > > > Thanks,
> > > > Andrii Biletskyi
> > > >
> > > > On Tue, Apr 7, 2015 at 4:11 PM, Andrii Biletskyi <
> > > > andrii.bilets...@stealth.ly> wrote:
> > > >
> > > > > Hi all,
> > > > >
> > > > > I wasn't able to send email to our thread (it says we exceeded
> > message
> > > > > size limit :)).
> > > > > So I'm starting the new one.
> > > > >
> > > > >
> > > > > Jun,
> > > > >
> > > > > Thanks again for the review. Answering your comments:
> > > > >
> > > > > 201. I'm not sure I understand how can we evolve Cluster in
> backward
> > > > > compatible way. In my understanding topic configs are not returned
> > > > > currently -
> > > > > in TMR_V0. Thus we need to add new property in Cluster - smth like
> > > > > private final Map<String, List<ConfigEntry>> topicConfigs;
> > > > > Which affects Cluster constructor, which is used in
> > > MetadataResponse.java
> > > > > - not sure whether we can change Cluster this way so it's backward
> > > > > compatible,
> > > > > I suppose - no.
> > > > > Let me know if I'm missing something...
> > > > >
> > > > > 300. Hm, so you propose to give up ReassignPartition as a separate
> > > > command?
> > > > > That's interesting, let's discuss it today in detail.
> > > > > Two small points here: 1) afaik currently replica-assignment
> argument
> > > in
> > > > > alter-topic
> > > > > (from TopicCommand) doesn't reassign partitions, it lets users
> > specify
> > > > > replica
> > > > > assignment for newly added partition (AddPartitionsListener) 2)
> > > > > ReassignPartitions
> > > > > command involves a little bit more than just changing replica
> > > assignment
> > > > > in zk.
> > > > > People are struggling with partition reassignment so I think it's
> > good
> > > to
> > > > > have explicit
> > > > > request for it so we can handle it independently, also as mentioned
> > > > > earlier we'll
> > > > > probably add in future some better status check procedure for this
> > > > > long-running
> > > > > request.
> > > > >
> > > > > 301. Good point. We also agreed to use clientId as an identifier
> for
> > > the
> > > > > requester -
> > > > > whether it's a producer client or admin. I think we can go with
> -1/-1
> > > > > approach.
> > > > >
> > > > > 302. Again, as said above replica-assignment in alter-topic doesn't
> > > > change
> > > > > replica assignment of the existing partitions. But we can think
> about
> > > it
> > > > > in general -
> > > > > how can we change topic replication factor? The easy way - we don't
> > > need
> > > > > it,
> > > > > we can use reassign partitions. Not sure whether we want to add
> > special
> > > > > logic
> > > > > to treat this case...
> > > > >
> > > > > 303.1. Okay, sure, I'll generalize topicExists().
> > > > > 303.2. I think, yes, we need separate verify methods as a status
> > check
> > > > > procedure,
> > > > > because respective requests are long running, and CLI user
> > potentially
> > > > > will
> > > > > asynchronously call reassign-partitions, do other stuff (e.g.
> create
> > > > > topics) periodically
> > > > > checking status of the partition reassignment. Anyway we'll have to
> > > > > implement this logic
> > > > > because it's a criterion of the completed Future of the reassign
> > > > > partitions async
> > > > > call, we'll to have make those methods just public.
> > > > > 303.3. If preferredReplica returns Future<Map<String, Errors>> than
> > > what
> > > > > is an error
> > > > > in terms of preferred replica leader election? As I understand we
> can
> > > > only
> > > > > check
> > > > > whether it has succeeded (leader == AR.head) or not _yet_.
> > > > >
> > > > > 304.1. Sure, let's add timeout to reassign/preferred replica.
> > > > > 304.2. This can be finalized after we discuss 300.
> > > > >
> > > > > 305. Misprints - thanks, fixed.
> > > > >
> > > > > Thanks,
> > > > > Andrii Biletskyi
> > > > >
> > > >
> > >
> >
>
