I'm going to update the KIP with new interface StateRestoreNotification
containing two methods, startRestore and endRestore.
While naming is very similar to methods already proposed on the
StateRestoreListener, the intent of these methods is not for user
notification of restore status. Instead these new methods are for internal
use by the state store to perform any required setup and teardown work due
to a batch restoration process.
Here's one current use case: when using RocksDB we should optimize for a
bulk load by setting Options.prepareForBulkload().
1. If the database has already been opened, we'll need to close it, set
the "prepareForBulkload" and re-open the database.
2. Once the restore is completed we'll need to close and re-open the
database with the "prepareForBulkload" option turned off.
While we are mentioning the RocksDB use case above, the addition of this
interface is not specific to any specific implementation of a persistent
state store.
Additionally, a separate interface is needed so that any user can implement
the state restore notification feature regardless of the state restore
callback used.
I'll also remove the "getStateRestoreListener" method and stick with the
notion of a "global" restore listener for now.
On Mon, Jun 19, 2017 at 1:05 PM, Bill Bejeck <bbej...@gmail.com> wrote:
> Yes it is, more of an oversight on my part, I'll remove it from the KIP.
>
>
> On Mon, Jun 19, 2017 at 12:48 PM, Matthias J. Sax <matth...@confluent.io>
> wrote:
>
>> Hi,
>>
>> I thinks for now it's good enough to start with a single global restore
>> listener. We can incrementally improve this later on if required. Of
>> course, if it's easy to do right away we can also be more fine grained.
>> But for KTable, we might want to add this after getting rid of all the
>> overloads we have atm.
>>
>> One question: what is the purpose of parameter "endOffset" in
>> #onRestoreEnd() -- isn't this the same value as provided in
>> #onRestoreStart() ?
>>
>>
>> -Matthias
>>
>>
>>
>> On 6/15/17 6:18 AM, Bill Bejeck wrote:
>> > Thinking about the custom StateRestoreListener approach and having a get
>> > method on the interface will really only work for custom state stores.
>> >
>> > So we'll need to provide another way for users to set behavior with
>> > provided state stores. The only option that comes to mind now is also
>> > adding a parameter to the StateStoreSupplier.
>> >
>> >
>> > Bill
>> >
>> >
>> > On Wed, Jun 14, 2017 at 5:39 PM, Bill Bejeck <bbej...@gmail.com> wrote:
>> >
>> >> Guozhang,
>> >>
>> >> Thanks for the comments.
>> >>
>> >> 1. As for the granularity, I agree that having one global
>> >> StateRestoreListener could be restrictive. But I think it's important
>> to
>> >> have a "setStateRestoreListener" on KafkaStreams as this allows users
>> to
>> >> define an anonymous instance that has access to local scope for
>> reporting
>> >> purposes. This is a similar pattern we use for
>> >> KafkaStreams.setStateListener.
>> >>
>> >> As an alternative, what if we add a method to the
>> BatchingStateRestoreCallback
>> >> interface named "getStateStoreListener". Then in an abstract adapter
>> >> class we return null from getStateStoreListener. But if users want to
>> >> supply a different StateRestoreListener strategy per callback they
>> would
>> >> simply override the method to return an actual instance.
>> >>
>> >> WDYT?
>> >>
>> >> 2. I'll make the required updates to pass in the ending offset at the
>> >> start as well as the actual name of the state store.
>> >>
>> >> Bill
>> >>
>> >>
>> >> On Wed, Jun 14, 2017 at 3:53 PM, Guozhang Wang <wangg...@gmail.com>
>> wrote:
>> >>
>> >>> Thanks Bill for the updated wiki. I have a couple of more comments:
>> >>>
>> >>> 1. Setting StateRestoreListener on the KafkaStreams granularity may
>> not be
>> >>> sufficient, as in the listener callback we do not which store it is
>> >>> restoring right now: if the topic is a changelog topic then from the
>> >>> `TopicPartition` we may be able to infer the state store name, but if
>> the
>> >>> topic is the source topic read as a KTable then we may not know which
>> >>> store
>> >>> it is restoring right now; plus forcing users to infer the state store
>> >>> name
>> >>> from the topic partition name would not be intuitive as well. Plus for
>> >>> different stores the listener may be implemented differently, and
>> setting
>> >>> a
>> >>> global listener would force users to branch on the topic-partition
>> names,
>> >>> similarly to what we did in the global timestamp extractor. On the
>> other
>> >>> hand, I also agree that setting the listener on the per-store
>> granularity
>> >>> may be a bit cumbersome since if users want to override it on a
>> specific
>> >>> store it needs to expose some APIs maybe at StateStoreSupplier. So
>> would
>> >>> love to hear other people's opinions.
>> >>>
>> >>> If we think that different implemented restoring callback may be less
>> >>> common, then I'd suggest at least replace the `TopicPartition`
>> parameter
>> >>> with the `String` store name and the `TaskId`?
>> >>>
>> >>> 2. I think we can pass in the `long endOffset` in the `onRestoreStart`
>> >>> function as well, as we will have read the endOffset already by then;
>> >>> otherwise users can still not be able to track the restoration
>> progress
>> >>> (e.g. how much percentage I have been restoring so far, to estimate
>> on how
>> >>> long I still need to wait).
>> >>>
>> >>>
>> >>> Guozhang
>> >>>
>> >>>
>> >>>
>> >>> On Wed, Jun 14, 2017 at 12:25 PM, Bill Bejeck <bbej...@gmail.com>
>> wrote:
>> >>>
>> >>>> Eno,
>> >>>>
>> >>>> Thanks for the comments.
>> >>>>
>> >>>> 1. As for having both restore and restoreAll, I kept the restore
>> method
>> >>> for
>> >>>> backward compatibility as that is what is used by current
>> implementing
>> >>>> classes. However as I think about it makes things cleaner to have a
>> >>> single
>> >>>> restore method taking a collection. I'll wait for others to weigh in,
>> >>> but
>> >>>> I'm leaning towards having a single restore method.
>> >>>>
>> >>>> 2. The "onBatchRestored" method is for keeping track of the restore
>> >>> process
>> >>>> as we load records from each poll request.
>> >>>>
>> >>>> For example if the change log contained 5000 records and
>> >>>> MAX_POLL_RECORDS is set to 1000, the "onBatchRestored" method would
>> get
>> >>>> called 5 times each time with the ending offset of the last record in
>> >>> the
>> >>>> batch and the count of the batch. I'll update the KIP to add
>> >>> comments
>> >>>> above the interface methods.
>> >>>>
>> >>>>
>> >>>> Thanks,
>> >>>> Bill
>> >>>>
>> >>>>
>> >>>> On Wed, Jun 14, 2017 at 11:49 AM, Eno Thereska <
>> eno.there...@gmail.com>
>> >>>> wrote:
>> >>>>
>> >>>>> Thanks Bill,
>> >>>>>
>> >>>>> A couple of questions:
>> >>>>>
>> >>>>>
>> >>>> 1. why do we need both restore and restoreAll, why can't we just have
>> >>> one,
>> >>>>> that takes a collection (i.e., restore all)? Are there cases when
>> >>> people
>> >>>>> want to restore one at a time? In that case, they could still use
>> >>>>> restoreAll with just 1 record in the collection right?
>> >>>>>
>> >>>>> 2. I don't quite get "onBatchRestored". Could you put a small
>> comment
>> >>> on
>> >>>>> top of all three methods. An example might help here.
>> >>>>>
>> >>>>> Thanks
>> >>>>> Eno
>> >>>>>
>> >>>>>
>> >>>>>> On 8 Jun 2017, at 18:05, Bill Bejeck <bbej...@gmail.com> wrote:
>> >>>>>>
>> >>>>>> Guozhang, Damian thanks for the comments.
>> >>>>>>
>> >>>>>> Giving developers the ability to hook into StateStore recovery
>> >>> phases
>> >>>> was
>> >>>>>> part of my original intent. However the state the KIP is in now
>> >>> won't
>> >>>>>> provide this functionality.
>> >>>>>>
>> >>>>>> As a result I'll be doing a significant revision of KIP-167. I'll
>> >>> be
>> >>>>> sure
>> >>>>>> to incorporate all your comments in the new revision.
>> >>>>>>
>> >>>>>> Thanks,
>> >>>>>> Bill
>> >>>>>>
>> >>>>>> On Wed, Jun 7, 2017 at 5:24 AM, Damian Guy <damian....@gmail.com>
>> >>>> wrote:
>> >>>>>>
>> >>>>>>> I'm largely in agreement with what Guozhang has suggested, i.e.,
>> >>>>>>> StateRestoreContext shouldn't have any setters on it and also need
>> >>> to
>> >>>>> have
>> >>>>>>> the end offset available such that people can use it derive
>> >>> progress.
>> >>>>>>> Slightly different, maybe the StateRestoreContext interface could
>> >>> be:
>> >>>>>>>
>> >>>>>>> long beginOffset()
>> >>>>>>> long endOffset()
>> >>>>>>> long currentOffset()
>> >>>>>>>
>> >>>>>>> One further thing, this currently doesn't provide developers the
>> >>>>> ability to
>> >>>>>>> hook into this information if they are using the built-in
>> >>> StateStores.
>> >>>>> Is
>> >>>>>>> this something we should be considering?
>> >>>>>>>
>> >>>>>>>
>> >>>>>>> On Tue, 6 Jun 2017 at 23:32 Guozhang Wang <wangg...@gmail.com>
>> >>> wrote:
>> >>>>>>>
>> >>>>>>>> Thanks for the updated KIP Bill, I have a couple of comments:
>> >>>>>>>>
>> >>>>>>>> 1) I'm assuming beginRestore / endRestore is called only once per
>> >>>> store
>> >>>>>>>> throughout the whole restoration process, and restoreAll is
>> called
>> >>>> per
>> >>>>>>>> batch. In that case I feel we can set the StateRestoreContext as
>> a
>> >>>>> second
>> >>>>>>>> parameter in restoreAll and in endRestore as well, and let the
>> >>>> library
>> >>>>> to
>> >>>>>>>> set the corresponding values instead and only let users to read
>> >>>> (since
>> >>>>>>> the
>> >>>>>>>> collection of key-value pairs do not contain offset information
>> >>>> anyways
>> >>>>>>>> users cannot really set the offset). The "lastOffsetRestored"
>> >>> would
>> >>>> be
>> >>>>>>> the
>> >>>>>>>> starting offset when called on `beginRestore`.
>> >>>>>>>>
>> >>>>>>>> 2) Users who wants to implement their own batch restoration
>> >>> callbacks
>> >>>>>>> would
>> >>>>>>>> now need to implement both `restore` and `restoreAll` while they
>> >>>> either
>> >>>>>>> let
>> >>>>>>>> `restoreAll` to call `restore` or implement the logic in
>> >>> `restoreAll`
>> >>>>>>> only
>> >>>>>>>> and never call `restore`. Maybe we can provide two abstract impl
>> >>> of
>> >>>>>>>> BatchingStateRestoreCallbacks which does beginRestore /
>> >>> endRestore as
>> >>>>>>>> no-ops, with one callback implementing `restoreAll` to call
>> >>> abstract
>> >>>>>>>> `restore` while the other implement `restore` to throw "not
>> >>> supported
>> >>>>>>>> exception" and keep `restoreAll` abstract.
>> >>>>>>>>
>> >>>>>>>> 3) I think we can also return the "offset limit" in
>> >>>>> StateRestoreContext,
>> >>>>>>>> which is important for users to track the restoration progress
>> >>> since
>> >>>>>>>> otherwise they could not tell how many percent of restoration has
>> >>>>>>>> completed. I.e.:
>> >>>>>>>>
>> >>>>>>>> public interface BatchingStateRestoreCallback extends
>> >>>>>>> StateRestoreCallback
>> >>>>>>>> {
>> >>>>>>>>
>> >>>>>>>> void restoreAll(Collection<KeyValue<byte[], byte []>> records,
>> >>>>>>>> StateRestoreContext
>> >>>>>>>> restoreContext);
>> >>>>>>>>
>> >>>>>>>> void beginRestore(StateRestoreContext restoreContext);
>> >>>>>>>>
>> >>>>>>>> void endRestore(StateRestoreContext restoreContext);
>> >>>>>>>> }
>> >>>>>>>>
>> >>>>>>>> public interface StateRestoreContext {
>> >>>>>>>>
>> >>>>>>>> long lastOffsetRestored();
>> >>>>>>>>
>> >>>>>>>> long endOffsetToRestore();
>> >>>>>>>>
>> >>>>>>>> int numberRestored();
>> >>>>>>>> }
>> >>>>>>>>
>> >>>>>>>>
>> >>>>>>>> Guozhang
>> >>>>>>>>
>> >>>>>>>>
>> >>>>>>>>
>> >>>>>>>> On Fri, Jun 2, 2017 at 9:16 AM, Bill Bejeck <bbej...@gmail.com>
>> >>>> wrote:
>> >>>>>>>>
>> >>>>>>>>> Guozhang, Matthias,
>> >>>>>>>>>
>> >>>>>>>>> Thanks for the comments. I have updated the KIP, (JIRA title
>> and
>> >>>>>>>>> description as well).
>> >>>>>>>>>
>> >>>>>>>>> I had thought about introducing a separate interface altogether,
>> >>> but
>> >>>>>>>>> extending the current one makes more sense.
>> >>>>>>>>>
>> >>>>>>>>> As for intermediate callbacks based on time or number of
>> >>> records, I
>> >>>>>>> think
>> >>>>>>>>> the latest update to the KIP addresses this point of querying
>> for
>> >>>>>>>>> intermediate results, but it would be per batch restored.
>> >>>>>>>>>
>> >>>>>>>>> Thanks,
>> >>>>>>>>> Bill
>> >>>>>>>>>
>> >>>>>>>>>
>> >>>>>>>>>
>> >>>>>>>>>
>> >>>>>>>>>
>> >>>>>>>>> On Fri, Jun 2, 2017 at 8:36 AM, Jim Jagielski <j...@jagunet.com>
>> >>>>> wrote:
>> >>>>>>>>>
>> >>>>>>>>>>
>> >>>>>>>>>>> On Jun 2, 2017, at 12:54 AM, Matthias J. Sax <
>> >>>>>>> matth...@confluent.io>
>> >>>>>>>>>> wrote:
>> >>>>>>>>>>>
>> >>>>>>>>>>> With regard to backward compatibility, we should not change
>> the
>> >>>>>>>> current
>> >>>>>>>>>>> interface, but add a new interface that extends the current
>> >>> one.
>> >>>>>>>>>>>
>> >>>>>>>>>>
>> >>>>>>>>>> ++1
>> >>>>>>>>>>
>> >>>>>>>>>>
>> >>>>>>>>>
>> >>>>>>>>
>> >>>>>>>>
>> >>>>>>>>
>> >>>>>>>> --
>> >>>>>>>> -- Guozhang
>> >>>>>>>>
>> >>>>>>>
>> >>>>>
>> >>>>>
>> >>>>
>> >>>
>> >>>
>> >>>
>> >>> --
>> >>> -- Guozhang
>> >>>
>> >>
>> >>
>> >
>>
>>
>
