Hi, Jinrui
> (Jinrui): Just like how I believe that multi-tenancy and multi-user are not > the same thing, regarding the suggestion to interchange "multi-tenancy" and > "resource control," while they may have some similarities in concept, they > are not interchangeable. First of all, you misunderstood me. I never said that multi-tenancy and multi-user are the same thing. What I want to emphasize is: one user per tenant is also one of the ways to realize multi-tenancy, and one tenant manages one database, which is also one of the ways to realize multi-tenancy, and dories, spanner also has such an implementation. > (Jinrui): multi-tenancy is a specific approach to architecture that enables > multiple tenants to share the same resources while keeping their data > isolated, whereas resource control is a broader term that can refer to any > mechanism used to manage and allocate system resources. Second: why can't we call it multi-tenant? We can achieve logical isolation of data through this feature and auth permission, and we can also manage and control users and database resources. Doesn't this just verify what you said above? Thanks, --------------------------------------- Houliang Qi BONC, Ltd ---- Replied Message ---- | From | Jinrui 张金瑞<329920...@qq.com.INVALID> | | Date | 04/12/2023 14:55 | | To | <dev@iotdb.apache.org> | | Subject | Re: [discuss] consider revert the feature of multi-tenancy | Hi, the following description assumes that multi-tenancy = resource control, the two words can be interchanged. Are there any objections? (Jinrui): Just like how I believe that multi-tenancy and multi-user are not the same thing, regarding the suggestion to interchange "multi-tenancy" and "resource control," while they may have some similarities in concept, they are not interchangeable. Multi-tenancy is a specific approach to architecture that enables multiple tenants to share the same resources while keeping their data isolated, whereas resource control is a broader term that can refer to any mechanism used to manage and allocate system resources. Based on this, continue to infer, if it is found that the released version has a bug a few days after it was released, should the release be cancelled, and it is better to wait for the bug to be fixed. (Jinrui): Actually, we were not discussing how to do code control. But since you brought up the issue of code quality, do you think the current code is of sufficient quality to be merged? Or did you notice the issue when reviewing the code? Of course, I understand that we don't have a quantifiable standard to judge whether code is mergeable or not, and everyone has their own criteria for making that judgment. Regarding your question about cancelling a release if a bug is found a few days after it was released, I think it depends on the severity of the bug and the impact it has on users. If the bug is minor and does not affect the core functionality of the software, it may be acceptable to wait for the bug to be fixed in the next release. However, if the bug is severe and causes major issues for users, it may be necessary to cancel the release and fix the bug immediately. As for the fact that it hasn’t been fixed for two days, this function is being discussed whether it should be discarded or not. Does it still take time to fix bugs? (Jinrui): I didn't fully understand what you meant, could you please explain it again? Thanks, Zhang Jinrui 2023年4月12日 下午1:35，Chao Wang <ccgow...@163.com> 写道： Hi all, It seems that Houliang didn't make it very clear before. Let me add some more information. If you think this code is doing multi tenant things now, why do we need to change the name to some other words like resource control ? Is it > just to prevent users from misunderstandings from the definition? If that's the case, does it mean that users perceive multi tenancy as different from what we do? This is contradictory. Why change multi-tenancy to resource control is just because everyone’s perception of multi-tenancy is not very unified. It may be more unified to change it to resource control. This should not be contradictory. Suppose, under the advance statement in pr, the following description assumes that multi-tenancy = resource control, the two words can be interchanged. Are there any objections? It has been two days since this code was merged, and I don't know if anyone is fixing the issue I mentioned. Bug is definitely accepted, but I don't think this issue belongs to a 'bug' because it failed even the most basic functional testing. From your test, it is true that there is a problem with this pr, and there is no pr fix at present. For the stability of the code base, it should be better to revert, even if this function is turned off by default (reminds me of developing a new version of distributed Framework, it seems that many codes are merged when they can't run, I don't know if it's suitable for this scenario). Then according to this inference, whether all PRs in the future have to repair and submit bugfix immediately once a bug is found, and it is better to revert after a certain period of time (assuming 2 days). Based on this, continue to infer, if it is found that the released version has a bug a few days after it was released, should the release be cancelled, and it is better to wait for the bug to be fixed. As for the fact that it hasn’t been fixed for two days, this function is being discussed whether it should be discarded or not. Does it still take time to fix bugs? Thanks! Chao Wang BONC ltd On 4/12/2023 11:26，张金瑞<329920...@qq.com.INVALID> wrote： Hi all, Although it seems that we have reached an agreement on what this PR really did, there are a few issues that I think are still unclear and need to be further discussed. > @Houliang said in previous mail: "a user is a tenant, and each tenant has different resources. This is also multi-tenancy" (Jinrui) Everyone difinitely can have their own understanding of multi-tenancy but I don't think "Tenant is equal to User". We can refer the definition of MultiTenancy from wikipedia here https://en.wikipedia.org/wiki/Multitenancy. I think no matter how we define the concept of multi tenant ourselves, the reaction of most users when they see this word is more important when they use IoTDB. On the other hand, If you think this code is doing multi tenant things now, why do we need to change the name to some other words like resource control ? Is it just to prevent users from misunderstandings from the definition? If that's the case, does it mean that users perceive multi tenancy as different from what we do? This is contradictory. > @Houliang said in previous mail: "I STRONGLY think that this PR does not violate the positioning and future development of IOTDB, so I STRONGLY think that revert is not needed" (Jinrui) I think the first part of my previous mail is not noticed. Maybe I need to emphasize it again. I didn't say that it was a MUST to perform a revert. What actually I said is if there is significant uncertainty in the code, revert is a quick way to make the repo keep stable. And I also appended a simple test report in last mail. And the test report showed that a very simple scenario cannot passed when enable the feature. It has been two days since this code was merged, and I don't know if anyone is fixing the issue I mentioned. Bug is definitely accepted, but I don't think this issue belongs to a 'bug' because it failed even the most basic functional testing. Thanks, Zhang Jinrui Original From:"Jialin Qiao"< qiaojia...@apache.org >; Date:2023/4/11 21:47 To:"dev"< dev@iotdb.apache.org >; Subject:Re: [discuss] consider revert the feature of multi-tenancy Hi, > I think we have reached a consensus from the discussion, change the name of this feature to resource control, and continue to contribute to this feature. +1, Thanks for Houliang and Yuhua's contribution! We indeed need a resource control module to keep the system safe :) Best, ————————————————— Jialin Qiao Apache IoTDB PMC Houliang Qi 于2023年4月11日周二 16:02写道： > > Hi, all > > > I think we have reached a consensus from the discussion, change the name of this feature to resource control, and continue to contribute to this feature. > Thank you for your concern. > > > > > > Thanks, > --------------------------------------- > Houliang Qi > BONC, Ltd > > > ---- Replied Message ---- > | From | Xiangdong Huang | > | Date | 04/11/2023 15:39 | > | To | | > | Subject | Re: [discuss] consider revert the feature of multi-tenancy | > Hi Houliang, > > It makes no sense to refer Doris. Doris is not a lightweight db, and > edge side is never its goal. > > The topic of this discussion is whether to revert the feature of multi-tenancy. > > I wonder why you fall into these words.... I think I have mentioned at > least twice (or maybe 3 times) that Jialin's suggestion is fine for > me. > > Best, > ----------------------------------- > Xiangdong Huang > School of Software, Tsinghua University > > > Houliang Qi 于2023年4月11日周二 15:05写道： > > Hi Jinrui, > > (Jinrui) From my perspective, Multi-tenancy is different from resource-control and they are not the different term for same thing. According to our implementation, current feature focus on the resource control on users of one tenant rather than on different tenants. If we did not reflect the wording `multi-tenancy` in the code, why do we use it on user docs and PR's description ? > > Sorry, I am not agree with you, from my perspective, a user is a tenant, and each tenant has different resources. This is also multi-tenancy. Even each tenant can only have one db. In our current implementation, a user is a tenant. > For doris, they also mention multi-tenancy, but it is limited user resources.[1], the same as our current implementation. > For Spanner, a tenant can also have only one db. [2] > The reason why I think that both multi-tenancy and resource-control are suitable for us is that what we are currently doing is to limit the functions of users or db resources. > On this point, I agree with Wang Chao's point of view. > > As for whether the multi-tenant function you mentioned affects the positioning of IoTDB, I don't think it is accurate. I personally think that the multi-tenant function is a term for resource isolation technology and will not affect the positioning of IoTDB. I don't know how you define the multi-tenant function. If it refers to the connection with the billing system of the cloud service provider, it may be another form. This discussion will not continue to discuss multi-tenancy. > > > > (Jinrui) REVERT does not mean REJECT. It is only a quick way to keep the code more reliable before we reach the same page. And furthermore, I don't think it is harmful or discouraging and it is only a regular way we use to replace hot-fix. > (Jinrui) The reviewers may be confused by the PR's description and then focus on whether `multi-tenant` should be integrated in current development stage of IoTDB. > > The topic of this discussion is whether to revert the feature of multi-tenancy. I STRONGLY think that this PR does not violate the positioning and future development of IOTDB, so I STRONGLY think that revert is not needed, as this function is not enabled by default, and we are continuing Iterate and refine this feature. Before the actual release, it is necessary to consider some scenarios and do some testing. > > > > [1] https://doris.apache.org/docs/dev/admin-manual/multi-tenant/ > [2] https://cloud.google.com/solutions/implementing-multi-tenancy-cloud-spanner > > > Thanks, > --------------------------------------- > Houliang Qi > BONC, Ltd > > > ---- Replied Message ---- > | From | Chao Wang | > | Date | 04/11/2023 13:42 | > | To | dev@iotdb.apache.org | > | Subject | Re: Re：Re: [discuss] consider revert the feature of multi-tenancy | > Everyone's contribution counts. But what we are talking about is whether `multi-tenancy` is suitable for current IoTDB's development. > From my perspective, Multi-tenancy is different from resource-control and they are not the different term for same thing. According to our implementation, current feature focus on the resource control on users of one tenant rather than on different tenants. If we did not reflect the wording `multi-tenancy` in the code, why do we use it on user docs and PR's description ? > > > As I said before, the description is indeed not very clear, and the description can be modified as a resource control. So what's the point of wondering if this pr is a multi-tenant function? Even if it is a multi-tenant function, how will it affect the development of IoTDB? > > > REVERT does not mean REJECT. It is only a quick way to keep the code more reliable before we reach the same page. And furthermore, I don't think it is harmful or discouraging and it is only a regular way we use to replace hot-fix. > > > Yes, revert is a normal process, and PR also has some problems. Let's discuss the reason for reverting this PR. As Xiangdong said, this is a feature that will affect the positioning of IoTDB, so how does this feature affect the positioning of IoTDB? > > > Agree. But if we don't make it clear before PR merged, pushing forward the discussion is better than directly merging, from my side. > > > Agree. I remember sending an email to discuss it. Does that mean that every PR needs to send an email to discuss clearly? After all, pushing forward the discussion is better than directly merging. > > > > > Thanks! > > > Chao Wang > BONC ltd > On 4/11/2023 13:10，张金瑞<329920...@qq.com.INVALID> wrote： > (Sorry for the format issue in previous mail) > ====== > Hi, all > > I tried this feature locally according to the User Manual, and I am blocked at the beginning. > > > Firstly, I didn't found the parameters `quota_enable` and `rate_limiter_type` in iotdb-common.properties to enable this functionality. > I am not sure whether it is by design but it is not aligned with the user manual. And I have to add these two parameters into configuration file manually. > > > Then, I tried the function to limit devices regarding a database and it seems some basic functionality is unexpected. See the test step below: > 1. create a databse named `root.iov` and insert 5 devices into it. > 2. run the command "set space quota devices=4 on root.iov" to set the device limit to 4. It can be executed successfully. (UNEXPECTED) > 3. try to insert a new devices. > 4. try to use the command "set space quota devices=8 on root.iov" to increase the threshold of device count but failed. (UNEXPECTED) > 5. I created another database named `sg2` and tried to set quota limit to it but failed (UNEXPECTED) > 6. After this test, I tried another test with more simple scenario but still failed. > > > The detailed test steps can be found in this doc: https://apache-iotdb.feishu.cn/docx/IerZdPFHroEbRYxKvihcBpncnie > > My confidence in the completeness of testing regarding this feature has greatly decreased. And at least from my perspective, I cannot guarantee how much impact this feature will have on the user experience. > > > On the other hand, I also have some thoughts to share with you regarding the questions raised in the previous email. > > > >> Leaving aside this feature, has the PR of the big feature we merged in been discussed in detail? How to define detailed discussion? > (Jinrui) I think currently we are focusing on the side effects of this PR, whether the discussion is detailed depends on whether we have enough confidence of this feature's definition and implementation. > > > >> I think that we should discuss some of our discussion points clearly at the beginning of the design, instead of how to revert the PR after the PR is merged. I think there is a problem with this process. > (Jinrui) Agree. But if we don't make it clear before PR merged, pushing forward the discussion is better than directly merging, from my side. > > > >> Who can guarantee that there are no bugs and no problems in the developed functions, and we are all improving through continuous iteration > (Jinrui) Yes. But the developer should do enough tests including different scenarios to ensure this feature can work smoothly. > > > >> However, reverting will undoubtedly be harmful to the community, will discourage the enthusiasm of community participants, and is very unfriendly to community participants > (Jinrui) REVERT does not mean REJECT. It is only a quick way to keep the code more reliable before we reach the same page. And furthermore, I don't think it is harmful or discouraging and it is only a regular way we use to replace hot-fix. > > > >> If in doubt, I think it would be better to raise it as soon as possible, instead of waiting for others to finish their hard work before questioning. > (Jinrui) My words has no doubt and offense to anyone and it is only a discussion about this issue. Contribution is welcomed and talking is also welcomed. If we notice some potential issue which is worth to be talked about, any time is ok I think. > > > >> This discuss is not for getting "+1" or "-1" (though anyone can reply the vote..). I just want to discuss that do we REALLY consider and analyze the feature and the implementation carefully? > (Jinrui) The reviewers may be confused by the PR's description and then focus on whether `multi-tenant` should be integrated in current development stage of IoTDB. > > > >> As for the name of this feature, in doris, it is called multi-tenancy[1], in hbase it is called quota[2], we can call it resource-control, I think it is ok. > (Jinrui) From my perspective, Multi-tenancy is different from resource-control and they are not the different term for same thing. According to our implementation, current feature focus on the resource control on users of one tenant rather than on different tenants. If we did not reflect the wording `multi-tenancy` in the code, why do we use it on user docs and PR's description ? > > > >> Another point is that the multi-tenancy function may be a function required by other companies' IOTDB releases, but will other people's contributions to the community affect the development of the community? > (Jinrui) Everyone's contribution counts. But what we are talking about is whether `multi-tenancy` is suitable for current IoTDB's development. > > > > > Thanks, > Zhang Jinrui > > > > > > > > > Original > > > > From:"Xiangdong Huang"< saint...@gmail.com >; > > Date:2023/4/11 12:27 > > To:"dev"< dev@iotdb.apache.org >; > > Subject:Re: [discuss] consider revert the feature of multi-tenancy > > > Hi Houliang, > > Notice that I never said the feature should be reverted because of > bugs.. The key point is the feature is harmful for Industry users > because most of them do not like cloud. (that is why I opt for > Jialin's suggestion). > > > I think that we should discuss some of our discussion points clearly at the beginning of the design, instead of how to revert the PR after the PR is merged. I think there is a problem with this process. > > It is of course right, but it does not mean that we can not revert a > PR if it is merged. > > > Leaving aside this feature, has the PR of the big feature we merged in been discussed in detail? How to define detailed discussion? > > Yes for each big feature we need a discussion in detail. As I have no > much time to join all the features, being the PMC chair, at least I > need to keep the project following its original destination or new > destination if we all agree. > > Considering my personal time, I judge and intervene featuers which may > change the product's position. That is why I spent time to discuss > whether we redesign the cluster mode, whether we split an IoTDB > instance into two (CN and DN), and whether we tell IoTDB is for > cloud-native... And that is why I do not care about more detailed > features.. > > Best, > ----------------------------------- > Xiangdong Huang > School of Software, Tsinghua University > > Houliang Qi 于2023年4月11日周二 09:51写道： > > > > Hi, all > > > > > > Leaving aside this feature, has the PR of the big feature we merged in been discussed in detail? How to define detailed discussion? > > > > I think that we should discuss some of our discussion points clearly at the beginning of the design, instead of how to revert the PR after the PR is merged. I think there is a problem with this process. > > > > Who can guarantee that there are no bugs and no problems in the developed functions, and we are all improving through continuous iteration. And this feature also refers to the design of some other excellent projects, such as doris and hbase. > > > > As for the name of this feature, in doris, it is called multi-tenancy[1], in hbase it is called quota[2], we can call it resource-control, I think it is ok. After all, we did not reflect the wording of multi-tenancy in the code implementation. > > > > > > > > [1] https://doris.apache.org/docs/dev/admin-manual/multi-tenant > > [2] https://hbase.apache.org/book.html#quota > > > > > > > > > > Thanks, > > --------------------------------------- > > Houliang Qi > > BONC, Ltd > > > > > > ---- Replied Message ---- > > | From | Chao Wang | > > | Date | 04/11/2023 09:15 | > > | To | dev@iotdb.apache.org | > > | Cc | dev@iotdb.apache.org | > > | Subject | Re: [discuss] consider revert the feature of multi-tenancy | > > Hi, Xiangdong, > > > > > > what is the side effect when we manually create a time series? > > > > > > How about the pr https://github.com/apache/iotdb/pull/9430, limit the timeseries number of cluster, anyone analyze the side effect about creating a time series? > > > > > > This discuss is not for getting "+1" or "-1" (though anyone can reply > > the vote..). > > I just want to discuss that do we REALLY consider and analyze the > > feature and the implementation carefully? > > > > > > Why not discuss before the PR submission, but wait until the PR submission before discussing, wouldn't it waste the energy of community participants? I have also seen emails sent before, not without notifying everyone. > > > > > > > > > > In addition, I think Jialin's suggestion is more reasonable. The description of this function may not be particularly clear. It can be said in another way, such as resource control. However, reverting will undoubtedly be harmful to the community, will discourage the enthusiasm of community participants, and is very unfriendly to community participants. If in doubt, I think it would be better to raise it as soon as possible, instead of waiting for others to finish their hard work before questioning. > > > > > > Another point is that the multi-tenancy function may be a function required by other companies' IOTDB releases, but will other people's contributions to the community affect the development of the community? I think it will be more conducive to the development of community diversity. > > > > > > > > > > > > > > Thanks! > > > > > > Chao Wang > > BONC ltd > > ccgow...@163.com > > On 4/10/2023 23:45，Xiangdong Huang wrote： > > Besides the above, when we merge this pr, we posted the design in the feishu[4] and discussed it online as least two times, and emailed and discussed it with everyone[5], it has been passed 10 days. > > > > I think I know this and I have shown my concern about the possible > > harm of this featuer to IoTDB's edge mode... > > > > 1) how many side-effects the feature will bring; > > We have done some tests under[1], which says with 20 databases and 1 user when we set `quota_enable` to true to enable the multi-tenancy feature, the write performance is only slowed down 1.75%, the read latency has not much difference, we will do more tests to show the side-effects in the feature. > > > > The experiment is rather simple... > > When we really want to show the added codes having no side-effects, > > all the exepriemnt settings should follow a rule that how to fully > > expose the possible problems. > > > > For example, as mult-tenancy limits the available # of devices, > > timeseries, and the spaces of disk, it should have side-effect on > > create new device/timeseries, and writing new data. > > So, > > - what is the side effect when we manually create a time series? > > - what is the side effect when we use automatical creating a time series? > > - what is the side effect when we write new data? (as the data can be > > compressed when it is flushed on disk in async mode, how to check the > > disk space?). Besides, as it impaces each write operation, we need to > > focus on write operstions which's batchsize=1. > > > > This discuss is not for getting "+1" or "-1" (though anyone can reply > > the vote..). > > I just want to discuss that do we REALLY consider and analyze the > > feature and the implementation carefully? > > > > If not, then this big feature is not the time to be merged (and I will > > call a vote then), and then let's rethink it and make it really > > available together. > > If yes, we also need to rethink it and improve it for better performance. > > > > > > Best, > > ----------------------------------- > > Xiangdong Huang > > School of Software, Tsinghua University > > > > Chao Wang 于2023年4月10日周一 19:14写道： > > > > Agree with Houliang's opinion. > > > > > > Thanks! > > > > > > Chao Wang > > BONC ltd > > On 4/10/2023 19:01，Houliang Qi wrote： > > -1 > > > > First of all, thanks Xiangdong for pointing out IoTDB's Charter. > > > > "RESOLVED, that the Apache IoTDB Project be and hereby is > > responsible for the creation and maintenance of software > > related to an IoT native database with high performance > > for data management and analysis, on the edge and the cloud." > > > > As the charter post, IoTDB can be deployed in the cloud, this is why we deploy the multi-tenancy feature. > > > > The cloud can be a public or private cloud if we can deploy only one IoTDB cluster, and manage multi databases and users with different resources, which will simplify the maintenance. > > > > -> 1) how many side-effects the feature will bring; > > > > We have done some tests under[1], which says with 20 databases and 1 user when we set `quota_enable` to true to enable the multi-tenancy feature, the write performance is only slowed down 1.75%, the read latency has not much difference, we will do more tests to show the side-effects in the feature. > > > > -> 2) how to reduce the effect when IoTDB is deployed on the edge. > > > > We supply one switch about this feature, called `quota_enable`, by default this value is false, so it has no effect when IoTDB is deployed on the edge. > > This also answers Jinrui's doubt. > > > > -> 3) some checks failed on WinOS, are they irrelevant? > > > > No, I think they are not irrelevant, the false check message is about the Compaction module, and > > I see the former pr[2][3] which have been merged 4 days ago has the same issue, so I suspect that the compaction module has occasional bugs > > > > -> 4) The feature SHOULD be discussed carefully in the community, rather that submit PRs and merged after some reviews. > > > > Besides the above, when we merge this pr, we posted the design in the feishu[4] and discussed it online as least two times, and emailed and discussed it with everyone[5], it has been passed 10 days. > > > > > > The IoTDB community is open and different opinions are welcome. After all, we all have the same original intention of wanting IoTDB's features to be more diverse. > > > > [1] https://apache-iotdb.feishu.cn/docx/DbqCd8t3EoxlCFx1yYicd9N4n4s > > [2] https://github.com/apache/iotdb/actions/runs/4625220921/jobs/8181102446 > > [3] https://github.com/apache/iotdb/actions/runs/4531046594/jobs/7980725316 > > [4] https://apache-iotdb.feishu.cn/docx/doxcnKOYKDmJ40FpVnVsPMd3nTg > > [5] https://lists.apache.org/thread/y6dqcm2o7qk0nbkllb61bp8cv6d3m1h7 > > > > > > > > > > > > Thanks, > > --------------------------------------- > > Houliang Qi > > BONC, Ltd > > > > > > ---- Replied Message ---- > > | From | 张金瑞<329920...@qq.com.INVALID> | > > | Date | 04/10/2023 15:03 | > > | To | dev | > > | Subject | Re：[discuss] consider revert the feature of multi-tenancy | > > +1, > > > > > > Agree with Xiangdong's opinion. > > And on the other hand, checking this PR's side effects may take lot of time and during this period, there may be lots of users using latest code to deploy/upgrade their systems. So the best practice is reverting this PR until the side-effect is eliminated > > > > > > > > Thanks, > > Zhang Jinrui，Apache IoTDB PMC > > > > > > > > Original > > > > > > > > From:"Xiangdong Huang"< saint...@gmail.com >; > > > > Date:2023/4/10 10:05 > > > > To:"dev"< dev@iotdb.apache.org >; > > > > Subject:[discuss] consider revert the feature of multi-tenancy > > > > > > Hi all, > > > > I see the multi-tenancy feature is merged, and several committers made > > a lot of contributions on that. > > > > As multi-tenancy is quite a big feature, which may change IoTDB's > > position. The feature SHOULD be discussed carefully in the community, > > rather that submit PRs and merged after some reviews. > > > > Therefore, I call to revert the PR and discuss ASAP about the feature > > after that. > > > > At least, the proposer need to answer the following questions, > > 1) how many side-effect the feature will bring; > > 2) how to reduce the effect when IoTDB is deployed on the edge. > > 3) some checks failed on WinOS, are they irrelevant? > > > > I don't mean of rejecting any big contribution to IoTDB or harming the > > community's diversity, but accepting this feature is really big > > decision and it deserves us to take time to deliberate. > > > > > > Attached IoTDB's Charter: > > "RESOLVED, that the Apache IoTDB Project be and hereby is > > responsible for the creation and maintenance of software > > related to an IoT native database with high performance > > for data management and analysis, on the edge and the cloud." > > > > > > [1] https://github.com/apache/iotdb/pull/9534/checks > > > > Best, > > ----------------------------------- > > Xiangdong Huang > > School of Software, Tsinghua University
