If we use compact topic, application first query request will lead to broker read all the topic message, it's a big impact to broker disk io, of course we can optimize this process like create a index for the keys, nevertheless if the key does not exist, broker still need to read all the index for the key to test if the key is exist. but all the things rocksdb or other kv components have done for us, so using mature kv components may be a better choice. Just like you said, the simpler the better for a message queue.
vongosling <[email protected]> 于2021年9月24日周五 下午3:58写道： > I just wonder if we could not open a new kV module, we really could not > solve the question you're mentioned? we could introduce a compression topic > for the last effective replay. Do we know how hard it is to maintain a > RocksDB, which my Facebook friends have talked to me a lot about, > their various kV production around RocksDB? > > Since we open-source RocketMQ, I've tried to help keep the architecture as > simple as possible, and I've warned you to learn to subtract while adding > things. While the code scale is still under control, I think the first > thing to really do is to design a good pluggable system that replaces the > existing simple SPI and API approach, but unfortunately, I haven't seen any > progress in that regard so far. > > Therefore, I think we should carefully examine our technical solutions. We > should not let the technical solutions just only for streams :-) > > heng du <[email protected]> 于2021年9月24日周五 下午12:43写道： > > > Totally agree with this proposal. The KV semantic storage can not only > > provide better support for streaming and connect, especially the storage > of > > checkpoints but also can be used to better manage metadata in the future. > > At the same time, compared to the compact topic, this proposal can > > significantly reduce user replay costs and save failure recovery time, > and > > KV semantic storage can actually be regarded as another index similar to > > CQ, which can be loaded on demand. In addition, there seems to be an > > unvoted RIP-22 proposal, but please don’t care :) > > > > > > vongosling <[email protected]> 于2021年9月23日周四 下午8:35写道： > > > > > Thanks for your clarify. I have been confused by RIP 22, It seems we > have > > > occupied 22, right[1]? > > > > > > > > > [1] https://github.com/apache/rocketmq/issues/2937 > > > > > > Amber Liu <[email protected]> 于2021年9月23日周四 下午3:29写道： > > > > > > > Sorry about the format problem, below is the correct one. > > > > RIP-22 Support KV semantic storageStatus > > > > > > > > - Current Status: Draft > > > > - Authors: ltamber <https://github.com/ltamber> > > > > > > > > > > > > - Shepherds: duhengforever <[email protected]> > > > > - Mailing List discussion: [email protected] > > > > > > > > > > > > - Pull Request: #PR_NUMBER > > > > - Released: <released_version> > > > > > > > > Background & Motivationwhat do we need to do > > > > > > > > - will we add a new module? *no*. > > > > - will we add new APIs? *yes*. > > > > > > > > > > > > - will we add new feature? *yes*. > > > > > > > > Why should we do that > > > > > > > > - Are there any problems of our current project? > > > > Currently, we can't get/put key-value from/into rocketmq, so if > we > > > use > > > > connector <https://github.com/apache/rocketmq-externals>, like > > > > FileSource, BinlogSource, we can't persist current read > > position/dump > > > > position to rocketmq rather than an external meta store like > > > > zookeeper/mysql, this will bring more operator risk by introduce > > > another > > > > component. this issue was also in streaming > > > > <https://github.com/apache/rocketmq-streams> scenarios when > > developer > > > > want to persist meta info like checkpoint. > > > > - What can we benefit proposed changes? > > > > rocketmq would not rely on external componet such as > > zookeeper/etcd > > > > to support meta data storage. > > > > > > > > Goals > > > > > > > > - What problem is this proposal designed to solve? > > > > Design a distribution persistent key-value store, application > > can > > > > put key-value into broker, and then get the value after a while, > in > > > the > > > > same time, it can also have the ability like compareAndSet, prefix > > get > > > > and > > > > so on. > > > > - To what degree should we solve the problem? > > > > This RIP must guarantee below point: > > > > 1. High availablity: if one broker in the broker group is down, > > > > application can put/get key-value through other broker, the > > > availablity > > > > is > > > > same with the message of rocketmq. > > > > 2. High capacity: the amount of key-value may very large, so > the > > > > key-value can not store in memory, we must store the key-value in > > > disk > > > > device. > > > > > > > > Non-Goals > > > > > > > > - What problem is this proposal NOT designed to solve? > > > > Nothing specific. > > > > - Are there any limits of this proposal? > > > > Nothing specific. > > > > > > > > ChangesArchitecture > > > > > > > > > > > > > > > > We will introduce rocksdb <https://github.com/facebook/rocksdb> to > > > persist > > > > key-value data, to say it more accurately, we use rocksdb to compact > > the > > > > value with the same key, we will not enable WAL in rocksdb to > decrease > > > > write amplification (most case), instead we can recover the rocksdb > > state > > > > and consistency by redo rocketmq commitlog. so the put/get flow > showed > > on > > > > the above figure is: > > > > put: the key-value message will put into commitlog first, and then > > > through > > > > the reputService redo commitlog, the key-value will put to rocksdb > > > > asynchronous, until this reput finished broker will not response to > > > client. > > > > get: application will get key-value from rocksdb thought broker > > directly. > > > > In addition, if we don't want introduce rocksdb > > > > <https://github.com/facebook/rocksdb> and the meta data content will > > not > > > > occupy too many memory, we can also use a key-value store base on > > memory > > > > map, there will a periodic serialization and persistence thread to > > > > guarantee data won't loss if broker restart or system abnormal > > shutdown, > > > > and the memory state consistency will also guaranteed by redo > rocketmq > > > > commitlog. > > > > Interface Design/Change > > > > > > > > - Method signature changes. *No* > > > > - Method behavior changes. *No* > > > > > > > > > > > > - CLI command changes. *No* > > > > - Log format or content changes. > > > > the properties of the message will add two flag, kv_opType > > indicate > > > > the request type is put key-value or get key-value, and key > indicate > > > the > > > > request key both in put or get operation. In order to pass the key > > > > through > > > > the network in the request header, we will encode/decode the > > key(byte > > > > array > > > > format) use base64 > > > > <https://docs.oracle.com/javase/8/docs/api/java/util/Base64.html> > > > > encoding method. > > > > > > > > > > > > Compatibility, Deprecation, and Migration Plan > > > > > > > > - Are backward and forward compatibility taken into consideration? > > > > New RequestCode between client and broker are added, so there > > are 2 > > > > compatibility situations: > > > > 1. old client+new broker: old clients won't make request with > > > > key-value flag, so broker will not receive key-value request, > which > > > keep > > > > all things as before. > > > > 2. new client+old broker: new clients will send key-value > > request, > > > > but the broker don't recognize the request code, and will return > > error > > > > msg. > > > > so we should upgrade broker first to support this feature. > > > > - Are there deprecated APIs? > > > > Nothing specific. > > > > > > > > > > > > - How do we do migration? > > > > Nothing specific. > > > > > > > > Implementation Outline > > > > > > > > We will implement the proposed changes by two phases. > > > > Phase 1 > > > > > > > > 1. Implement reput logic from commitlog to rocksdb. > > > > 2. Implement broker support key-value request and response. > > > > > > > > > > > > 1. Implement client support key-value request and response. > > > > 2. Implement key-value store use memory map. > > > > > > > > > > > > 1. Implement key-value store use rocksdb. > > > > > > > > Phase 2 > > > > > > > > 1. Implement prefix get semantics. > > > > 2. Implement compareAndSet semantics. > > > > > > > > > > > > 1. Implement rocksdb snapshot export/import. > > > > > > > > > > > > Amber Liu <[email protected]> 于2021年9月23日周四 上午10:10写道： > > > > > > > > > # RIP-22 Support KV semantic storage > > > > > > > > > > ## Status > > > > > - Current Status: Draft > > > > > - Authors: [ltamber](https://github.com/ltamber) > > > > > - Shepherds: [duhengforever](mailto:[email protected]) > > > > > - Mailing List discussion: <[email protected]> > > > > > - Pull Request: #PR_NUMBER > > > > > - Released: <released_version> > > > > > ## Background & Motivation > > > > > ### what do we need to do > > > > > - will we add a new module? **no**. > > > > > - will we add new APIs? **yes**. > > > > > - will we add new feature? **yes**. > > > > > ### Why should we do that > > > > > - Are there any problems of our current project? > > > > > Currently, we can't get/put key-value from/into rocketmq, so if > we > > > use > > > > > [connector](https://github.com/apache/rocketmq-externals), like > > > > > FileSource, BinlogSource, we can't persist current read > position/dump > > > > > position to rocketmq rather than an external meta store like > > > > > zookeeper/mysql, this will bring more operator risk by introduce > > > another > > > > > component. this issue was also in [streaming]( > > > > > https://github.com/apache/rocketmq-streams) scenarios when > developer > > > > want > > > > > to persist meta info like checkpoint. > > > > > - What can we benefit proposed changes? > > > > > rocketmq would not rely on external componet such as > > zookeeper/etcd > > > to > > > > > support meta data storage. > > > > > ### Goals > > > > > - What problem is this proposal designed to solve? > > > > > Design a distribution persistent key-value store, application > can > > > put > > > > > key-value into broker, and then get the value after a while, in the > > > same > > > > > time, it can also have the ability like compareAndSet, prefix get > and > > > so > > > > on. > > > > > - To what degree should we solve the problem? > > > > > This RIP must guarantee below point: > > > > > 1. High availablity: if one broker in the broker group is down, > > > > > application can put/get key-value through other broker, the > > availablity > > > > is > > > > > same with the message of rocketmq. > > > > > 2. High capacity: the amount of key-value may very large, so the > > > > > key-value can not store in memory, we must store the key-value in > > disk > > > > > device. > > > > > ### Non-Goals > > > > > - What problem is this proposal NOT designed to solve? > > > > > Nothing specific. > > > > > - Are there any limits of this proposal? > > > > > Nothing specific. > > > > > ## Changes > > > > > ### Architecture > > > > >  > > > > > We will introduce [rocksdb](https://github.com/facebook/rocksdb) > to > > > > > persist key-value data, to say it more accurately, we use rocksdb > to > > > > > compact the value with the same key, we will not enable WAL in > > rocksdb > > > to > > > > > decrease write amplification (most case), instead we can recover > the > > > > > rocksdb state and consistency by redo rocketmq commitlog. so the > > > put/get > > > > > flow showed on the above figure is: > > > > > put: the key-value message will put into commitlog first, and then > > > > through > > > > > the `reputService` redo commitlog, the key-value will put to > rocksdb > > > > > asynchronous, until this reput finished broker will not response to > > > > client. > > > > > get: application will get key-value from rocksdb thought broker > > > directly. > > > > > In addition, if we don't want introduce [rocksdb]( > > > > > https://github.com/facebook/rocksdb) and the meta data content > will > > > not > > > > > occupy too many memory, we can also use a key-value store base on > > > memory > > > > > map, there will a periodic serialization and persistence thread to > > > > > guarantee data won't loss if broker restart or system abnormal > > > shutdown, > > > > > and the memory state consistency will also guaranteed by redo > > rocketmq > > > > > commitlog. > > > > > ### Interface Design/Change > > > > > - Method signature changes. **No** > > > > > - Method behavior changes. **No** > > > > > - CLI command changes. **No** > > > > > - Log format or content changes. > > > > > the properties of the message will add two flag, `kv_opType` > > > indicate > > > > > the request type is put key-value or get key-value, and `key` > > indicate > > > > the > > > > > request key both in put or get operation. In order to pass the key > > > > through > > > > > the network in the request header, we will encode/decode the > key(byte > > > > array > > > > > format) use [base64]( > > > > > https://docs.oracle.com/javase/8/docs/api/java/util/Base64.html) > > > > > encoding method. > > > > >  > > > > > ### Compatibility, Deprecation, and Migration Plan > > > > > - Are backward and forward compatibility taken into consideration? > > > > > New RequestCode between client and broker are added, so there > are > > 2 > > > > > compatibility situations: > > > > > 1. old client+new broker: old clients won't make request with > > > > > key-value flag, so broker will not receive key-value request, which > > > keep > > > > > all things as before. > > > > > 2. new client+old broker: new clients will send key-value > > request, > > > > but > > > > > the broker don't recognize the request code, and will return error > > msg. > > > > so > > > > > we should upgrade broker first to support this feature. > > > > > - Are there deprecated APIs? > > > > > Nothing specific. > > > > > - How do we do migration? > > > > > Nothing specific. > > > > > ### Implementation Outline > > > > > We will implement the proposed changes by two phases. > > > > > #### Phase 1 > > > > > 1. Implement reput logic from commitlog to rocksdb. > > > > > 2. Implement broker support key-value request and response. > > > > > 3. Implement client support key-value request and response. > > > > > 4. Implement key-value store use memory map. > > > > > 5. Implement key-value store use rocksdb. > > > > > #### Phase 2 > > > > > 1. Implement prefix get semantics. > > > > > 2. Implement compareAndSet semantics. > > > > > 3. Implement rocksdb snapshot export/import. > > > > > > > > > > > > > > > > > > -- > > > Best Regards :-) > > > > > > > > -- > Best Regards :-) >
