> Users can use > ``` > MessageId.fromByteArrayWithTopic() > ``` > or > ``` > TopicMessageId.fromByteArray(); > TopicMessageId.create(); > ```
I think this is a good point. `TopicName` don't in the `MessageId` means that multiConsumer must know that this topic is a partitioned topic and needs to assign `TopicName` to `MessageId` and doesn't need to add new interfaces. Thanks, Bo PengHui Li <peng...@apache.org> 于2022年12月22日周四 09:50写道： > > > Because `TopicMessageId` is constructed by the Pulsar Client library > itself, which can guarantee `getOwnerTopic()` returns the correct > topic name. The benefit of passing a `TopicMessageId` rather than the > combination of a topic name and a `MessageId` is, users won't need to > care about how to get the correct topic name for a given partition by > themselves. > > This is a good point of why we should not introduce seek(TopicName, > MessageId) > It's better also to mention this part in the proposal. It will help us to > understand > why seek(TopicName, MessageId) is not a good way for Pulsar. > > Thanks, > Penghui > > On Thu, Dec 22, 2022 at 9:48 AM PengHui Li <peng...@apache.org> wrote: > > > > In short, `seek(msgId)` will call `seek(TopicMessageId)` if `msgId` is > > a `TopicMessageId`. > > > > Does it look like we don't need to add the following new APIs? > > > > ``` > > void seek(TopicMessageId topicMessageId) throws PulsarClientException; > > CompletableFuture<Void> seekAsync(TopicMessageId topicMessageId); > > ``` > > > > Users can use > > ``` > > MessageId.fromByteArrayWithTopic() > > ``` > > or > > ``` > > TopicMessageId.fromByteArray(); > > TopicMessageId.create(); > > ``` > > to construct a TopicMessageId instance. > > > > But we can use the existing seek API with the constructed TopicMessageId > > instance > > > > ``` > > consumer.seek(MessageId messageId); > > ``` > > > > Thanks, > > Penghui > > > > On Thu, Dec 22, 2022 at 12:35 AM Yunze Xu <y...@streamnative.io.invalid> > > wrote: > > > >> > but it needs to cast the `TopicMessageId` from `MessageId`, which is > >> very user-unfriendly. > >> > >> Sorry I think my proposal doesn't express it well. In my original > >> thought, no cast is needed, please see the update in > >> https://github.com/apache/pulsar/issues/18616. > >> > >> In short, `seek(msgId)` will call `seek(TopicMessageId)` if `msgId` is > >> a `TopicMessageId`. > >> > >> Thanks, > >> Yunze > >> > >> On Wed, Dec 21, 2022 at 11:26 PM 丛搏 <bog...@apache.org> wrote: > >> > > >> > > If you mean `msg.getTopicName()`, how can you declare it's better than > >> > > `msgId.getOwnerTopic()`? > >> > > >> > > It can. Because the `TopicMessageIdImpl` already contains the correct > >> > > topic. That's the point. > >> > > >> > ``` > >> > var msgId = (TopicMessageId) multiTopicsConsumer.getMessageId(); > >> > ``` > >> > if `msgId.getOwnerTopic()` is the interface of `MessageId`, I have no > >> > problem. but it needs to cast the `TopicMessageId` from `MessageId`, > >> > which is very user-unfriendly. And it doesn't make sense. > >> > > >> > > I don't know what you're thinking about using > >> > > `consumer.seek(msg.getTopicName(), msg.getMessageId()` for a single > >> > > topic consumer. If it's accepted, and you want to unify the use case > >> > > of `seek`, the original `seek` API should be deprecated and much > >> > > existing code could be affected. If it's not accepted, users have to > >> > > distinguish if a consumer is a multi-topics consumer. > >> > > >> > the same as `consumer.seek(TopicMessageId topicMessageId)` for a > >> > single topic consumer is also strange. My point is either TopicName > >> > belongs to MessageId or separate the two. it's not a good interface > >> > implementation to couple them together. Very unclear. > >> > > >> > Thanks, > >> > Bo > >> > > >> > Yunze Xu <y...@streamnative.io.invalid> 于2022年12月21日周三 22:46写道： > >> > > > >> > > > If messageID does not contain `TopicName`, the `TopicName` is best > >> get from msg. > >> > > > >> > > If you mean `msg.getTopicName()`, how can you declare it's better than > >> > > `msgId.getOwnerTopic()`? > >> > > > >> > > > but it still cannot avoid `TopicName` for marking this `MessageID` > >> belongs to this topic. > >> > > > >> > > It can. Because the `TopicMessageIdImpl` already contains the correct > >> > > topic. That's the point. > >> > > > >> > > > if using `TopicMessageId` also has the same problem, why we need to > >> use `TopicMessageId` not `MessageId` > >> > > > >> > > Because `TopicMessageId` is constructed by the Pulsar Client library > >> > > itself, which can guarantee `getOwnerTopic()` returns the correct > >> > > topic name. The benefit of passing a `TopicMessageId` rather than the > >> > > combination of a topic name and a `MessageId` is, users won't need to > >> > > care about how to get the correct topic name for a given partition by > >> > > themselves. > >> > > > >> > > The key point is that if there is only one valid value for an > >> > > argument, which relies on the other argument, then the API design is > >> > > bad. Assume you need to use the Pulsar client like: > >> > > > >> > > ``` > >> > > // numberOfMessages must be the same with msgIds.size(), otherwise, an > >> > > exception will be thrown > >> > > consumer.acknowledge(numberOfMessages, msgIds); > >> > > ``` > >> > > > >> > > With the API of this proposal, users don't need to care much about how > >> > > to call `seek` correctly, except the MessageId is returned by > >> > > Producer#send. `consumer.seek(msg.getMessageId())` works for all > >> > > cases. > >> > > > >> > > With the `seek(String, MessageId)` API, you have to write more > >> > > explanations like: > >> > > 1. If the consumer only subscribes to a topic, use > >> > > `consumer.seek(msg.getMessageId())`. > >> > > 2. If the consumer subscribes to multiple topics, use > >> > > `consumer.seek(topic, msg.getMessageId())`. The topic must be what the > >> > > message belongs to, so you have to use the correct topic like > >> > > `consumer.seek(msg.getTopicName(), msg.getMessageId()`. Otherwise, > >> > > seek would fail. > >> > > > >> > > I don't know what you're thinking about using > >> > > `consumer.seek(msg.getTopicName(), msg.getMessageId()` for a single > >> > > topic consumer. If it's accepted, and you want to unify the use case > >> > > of `seek`, the original `seek` API should be deprecated and much > >> > > existing code could be affected. If it's not accepted, users have to > >> > > distinguish if a consumer is a multi-topics consumer. > >> > > > >> > > Thanks, > >> > > Yunze > >> > > > >> > > On Wed, Dec 21, 2022 at 8:50 PM 丛搏 <bog...@apache.org> wrote: > >> > > > > >> > > > Hi, Yunze: > >> > > > > >> > > > < ```java > >> > > > < var msg = multiTopicsConsumer.receive(); > >> > > > < var msgId = (TopicMessageId) multiTopicsConsumer.getMessageId(); > >> > > > < consumer.seek(msgId.getOwnerTopic(), msgId); > >> > > > < ``` > >> > > > > >> > > > the code can be like this: > >> > > > ```java > >> > > > var msg = anyConsumer.receive(); > >> > > > var msgId = anyConsume.getMessageId(); > >> > > > consumer.seek(msg, msgId); > >> > > > ``` > >> > > > If messageID does not contain `TopicName`, the `TopicName` is best > >> get from msg. > >> > > > > >> > > > < What's different is that the offset in Kafka can represent a > >> position > >> > > > < of ANY partition, while the MessageId in Pulsar can only > >> represent the > >> > > > < position of A SPECIFIC partition. > >> > > > > >> > > > Although MessageId in Pulsar can only represent the position of A > >> > > > SPECIFIC partition, but it still needs a TopicName. `LedgerID` and > >> > > > `EntryID` do not mean that this `MessageID` belongs to a topic > >> > > > (although it does belong), but it still cannot avoid `TopicName` for > >> > > > marking this `MessageID` belongs to this topic. > >> > > > > >> > > > > And in Pulsar, we also do not expose > >> > > > > the partition concept, if we introduce the seek API with the topic > >> > > > > name as the argument, we have to explain in detail about what's > >> the > >> > > > > topic name for a partition. It could be a very confusing thing > >> from my > >> > > > > experience when I explained the "partition" concept in community. > >> > > > > >> > > > if using `TopicMessageId` also has the same problem, why we need to > >> > > > use `TopicMessageId` not `MessageId` > >> > > > > >> > > > Thanks, > >> > > > Bo > >> > > > > >> > > > Yunze Xu <y...@streamnative.io.invalid> 于2022年12月21日周三 16:59写道： > >> > > > > > >> > > > > Hi Bo, > >> > > > > > >> > > > > If we have the `seek` API that accepts a topic name, how to use > >> seek > >> > > > > for a single topic consumer and multi-topics consumer will be > >> > > > > different. > >> > > > > > >> > > > > ```java > >> > > > > var msg = singleTopicConsumer.receive(); > >> > > > > var msgId = singleTopicConsumer.getMessageId(); > >> > > > > consumer.seek(msgId); > >> > > > > ``` > >> > > > > > >> > > > > ```java > >> > > > > var msg = multiTopicsConsumer.receive(); > >> > > > > var msgId = (TopicMessageId) multiTopicsConsumer.getMessageId(); > >> > > > > consumer.seek(msgId.getOwnerTopic(), msgId); > >> > > > > ``` > >> > > > > > >> > > > > It's not as clear as you have thought. A question could come from > >> the > >> > > > > code above: since we can get the key (topic name) from `msgId` > >> itself, > >> > > > > why do we need another argument? > >> > > > > > >> > > > > What's worse is that users have to specify the correct topic > >> name. For > >> > > > > a partitioned topic, if users specified another partition, the > >> `seek` > >> > > > > operation would fail. If they specified something like > >> > > > > `multiTopicsConsumer.getTopic()`, it would also fail because other > >> > > > > APIs like `Consumer#getTopic()` doesn't return the correct topic > >> name. > >> > > > > > >> > > > > If there is only one correct topic name for a given > >> TopicMessageId, > >> > > > > what's the meaning of making it as a required argument? > >> > > > > > >> > > > > BTW, let's see Kafka client's commit API: > >> > > > > > >> > > > > ```java > >> > > > > public void commitSync(Map<TopicPartition,OffsetAndMetadata> > >> offsets) > >> > > > > ``` > >> > > > > > >> > > > > What's different is that the offset in Kafka can represent a > >> position > >> > > > > of ANY partition, while the MessageId in Pulsar can only > >> represent the > >> > > > > position of A SPECIFIC partition. And in Pulsar, we also do not > >> expose > >> > > > > the partition concept, if we introduce the seek API with the topic > >> > > > > name as the argument, we have to explain in detail about what's > >> the > >> > > > > topic name for a partition. It could be a very confusing thing > >> from my > >> > > > > experience when I explained the "partition" concept in community. > >> > > > > > >> > > > > Thanks, > >> > > > > Yunze > >> > > > > > >> > > > > > >> > > > > On Wed, Dec 21, 2022 at 3:20 PM 丛搏 <bog...@apache.org> wrote: > >> > > > > > > >> > > > > > Hi Yunze, > >> > > > > > > >> > > > > > add `TopicMessageId ` will couple messageID and `topic name` > >> together, > >> > > > > > which is very unclear for non-partition-topic. > >> > > > > > > >> > > > > > ``` > >> > > > > > void seek(String topicName, MessageId messageId) throws > >> PulsarClientException; > >> > > > > > List<Map<String, MessageId>> getLastTopicMessageId() throws > >> > > > > > PulsarClientException; > >> > > > > > ``` > >> > > > > > If the interface is designed in this way, it may be simpler, > >> easier to > >> > > > > > understand, and more intuitive for users, and MessageID will > >> not be > >> > > > > > coupled with TopicName. > >> > > > > > > >> > > > > > because this PIP has already initiated a VOTE, so I will sync > >> this > >> > > > > > reply to PIP-224-VOTE[0] > >> > > > > > > >> > > > > > Thanks, > >> > > > > > Bo > >> > > > > > [0] > >> https://lists.apache.org/thread/mbrpjsgrgwrlkdpvkk738jxnlk7rf4qk > >> > > > > > > >> > > > > > Yunze Xu <y...@streamnative.io.invalid> 于2022年12月9日周五 14:33写道： > >> > > > > > > > >> > > > > > > Hi Jiaqi, > >> > > > > > > > >> > > > > > > Let's move to > >> https://lists.apache.org/thread/mbrpjsgrgwrlkdpvkk738jxnlk7rf4qk > >> > > > > > > for the vote. > >> > > > > > > > >> > > > > > > Thanks, > >> > > > > > > Yunze > >> > > > > > > > >> > > > > > > On Fri, Dec 9, 2022 at 1:54 PM Jiaqi Shen < > >> gleiphir2...@gmail.com> wrote: > >> > > > > > > > > >> > > > > > > > This is make sense to me, +1 > >> > > > > > > > > >> > > > > > > > Thanks, > >> > > > > > > > Jiaqi Shen > >> > > > > > > > > >> > > > > > > > > >> > > > > > > > Yunze Xu <y...@streamnative.io.invalid> 于2022年12月7日周三 > >> 13:51写道： > >> > > > > > > > > >> > > > > > > > > Hi Baodi, > >> > > > > > > > > > >> > > > > > > > > I decided not to change the behavior of the > >> `negativeAcknowledge` > >> > > > > > > > > method. I just checked again that there is no exception > >> signature for > >> > > > > > > > > this method and there is no asynchronous version like > >> > > > > > > > > `negativeAcknowledgeAsync`. To keep the API compatible, > >> we should not > >> > > > > > > > > add an exception signature, which would be required if a > >> > > > > > > > > `PulsarClientException` was thrown. > >> > > > > > > > > > >> > > > > > > > > Thanks, > >> > > > > > > > > Yunze > >> > > > > > > > > > >> > > > > > > > > On Tue, Nov 29, 2022 at 10:12 PM Baodi Shi > >> <baodi....@icloud.com.invalid> > >> > > > > > > > > wrote: > >> > > > > > > > > > > >> > > > > > > > > > Hi, Yunze: > >> > > > > > > > > > > >> > > > > > > > > > Thanks for your proposal. That Looks good to me. > >> > > > > > > > > > > >> > > > > > > > > > `negativeAcknowledge` also needs to add the same checks > >> as the new > >> > > > > > > > > acknowledge interface. > >> > > > > > > > > > > >> > > > > > > > > > > This interface doesn't add any acknowledge overload > >> because the > >> > > > > > > > > overloads are already too many. But it will make the > >> behavior clear. > >> > > > > > > > > > I think since we exposed the TopicMessageId, it would > >> be better to add > >> > > > > > > > > overloaded interfaces (even if the overloads are a lot). > >> This can users to > >> > > > > > > > > clearly associate the use cases of MultiTopicConsumer and > >> TopicMessageId. > >> > > > > > > > > > > >> > > > > > > > > > Also, while it's okay to use TopicMessageId param on a > >> single consumer, > >> > > > > > > > > I guess we shouldn't allow users to use it. > >> > > > > > > > > > > >> > > > > > > > > > In this way, users are clearly aware that > >> TopicMessageId is used when > >> > > > > > > > > using MultiTopicConsumer and MessageId is used when using > >> > > > > > > > > SingleTopicConsumer.(Maybe it's not a good idea) > >> > > > > > > > > > > >> > > > > > > > > > > >> > > > > > > > > > Thanks, > >> > > > > > > > > > Baodi Shi > >> > > > > > > > > > > >> > > > > > > > > > > 2022年11月29日 15:57，Yunze Xu > >> <y...@streamnative.io.INVALID> 写道： > >> > > > > > > > > > > > >> > > > > > > > > > >> Is there a case where the user uses the messageId > >> returned by the > >> > > > > > > > > > > producer to seek in the consumer? Is this a good > >> behavior? > >> > > > > > > > > > > > >> > > > > > > > > > > Yes. I think it should be acceptable. To correct my > >> previous point, > >> > > > > > > > > > > now I think the MessageId returned by send should > >> also be able to be > >> > > > > > > > > > > applied for seek or acknowledge. > >> > > > > > > > > > > > >> > > > > > > > > > >> even with the > >> > > > > > > > > > > current proposal, it may return null when getting the > >> topic from > >> > > > > > > > > > > TopicMessageId for backward compatibility. > >> > > > > > > > > > > > >> > > > > > > > > > > No. It may return null just because Java doesn't > >> allow a non-null > >> > > > > > > > > > > returned value. The internal implementations of > >> > > > > > > > > > > TopicMessageId#getOwerTopic should return a non-null > >> topic name to > >> > > > > > > > > > > avoid null check. > >> > > > > > > > > > > > >> > > > > > > > > > > When I mentioned **the implementation of > >> getTopicName() must return > >> > > > > > > > > > > null**, the assumption is that MessageId#toByteArray > >> serializes the > >> > > > > > > > > > > topic name if adding the `getTopicName()` method. > >> However, in this > >> > > > > > > > > > > proposal, `TopicMessageId#toByteArray` won't. See the > >> implementation > >> > > > > > > > > > > of `TopicMessageId#create`. It's only a wrapper for > >> an arbitrary > >> > > > > > > > > > > MessageId implementation. > >> > > > > > > > > > > > >> > > > > > > > > > > Thanks, > >> > > > > > > > > > > Yunze > >> > > > > > > > > > > > >> > > > > > > > > > > On Tue, Nov 29, 2022 at 2:47 PM Zike Yang < > >> z...@apache.org> wrote: > >> > > > > > > > > > >> > >> > > > > > > > > > >> Hi Yunze, > >> > > > > > > > > > >> > >> > > > > > > > > > >> Thanks for your proposal. Quoted from your GitHub > >> comments[0]: > >> > > > > > > > > > >> > >> > > > > > > > > > >>> There is also a case when MessageId is returned > >> from Producer#send. > >> > > > > > > > > In this case, the returned MessageId should only used for > >> serialization > >> > > > > > > > > > >> > >> > > > > > > > > > >> Is there a case where the user uses the messageId > >> returned by the > >> > > > > > > > > > >> producer to seek in the consumer? Is this a good > >> behavior? > >> > > > > > > > > > >> > >> > > > > > > > > > >>> If we added the method directly to MessageId, to > >> keep the backward > >> > > > > > > > > compatibility, the implementation of getTopicName() must > >> return null, which > >> > > > > > > > > is not a good design. > >> > > > > > > > > > >> > >> > > > > > > > > > >> I think it's a trade-off. If I understand correctly, > >> even with the > >> > > > > > > > > > >> current proposal, it may return null when getting > >> the topic from > >> > > > > > > > > > >> TopicMessageId for backward compatibility. The > >> current > >> > > > > > > > > > >> TopicMessageIdImpl doesn't serialize the topic > >> information. > >> > > > > > > > > > >> > >> > > > > > > > > > >> > >> > > > > > > > > > >> [0] > >> > > > > > > > > > >> https://github.com/apache/pulsar/issues/18616#issuecomment-1328609346 > >> > > > > > > > > > >> > >> > > > > > > > > > >> Thanks, > >> > > > > > > > > > >> Zike Yang > >> > > > > > > > > > >> > >> > > > > > > > > > >> On Mon, Nov 28, 2022 at 12:22 PM Yunze Xu > >> > > > > > > > > <y...@streamnative.io.invalid> wrote: > >> > > > > > > > > > >>> > >> > > > > > > > > > >>> Hi all, > >> > > > > > > > > > >>> > >> > > > > > > > > > >>> I've opened a PIP to discuss: > >> > > > > > > > > https://github.com/apache/pulsar/issues/18616. > >> > > > > > > > > > >>> > >> > > > > > > > > > >>> The consumer's MessageId related APIs have some > >> hidden requirements > >> > > > > > > > > > >>> and flakiness and some behaviors are not documented > >> well. This > >> > > > > > > > > > >>> proposal will introduce a TopicMessageId interface > >> that exposes a > >> > > > > > > > > > >>> method to get a message's owner topic. > >> > > > > > > > > > >>> > >> > > > > > > > > > >>> P.S. There was an email [1] that didn't add the > >> "[DISCUSS]" label, > >> > > > > > > > > > >>> which might be a little confusing. So I sent the > >> email again for > >> > > > > > > > > > >>> discussion. Please do not reply to the previous > >> email. > >> > > > > > > > > > >>> > >> > > > > > > > > > >>> [1] > >> https://lists.apache.org/thread/6gj16pmrjk6ncsd30xrl20pr5ng6t61o > >> > > > > > > > > > >>> > >> > > > > > > > > > >>> Thanks, > >> > > > > > > > > > >>> Yunze > >> > > > > > > > > > > >> > > > > > > > > > >> > >
