Hi David, As explained in the motivation section of the KIP, the problem is that if log retention is too small, we may lose data; and if log retention is too large, then we waste disk space. Therefore, we need to solve one if the two problems -- allow data to be persisted longer for consumption if log retention is set too small, or allow data to be expired earlier if log retention is too large. I think the KIP probably needs to make this clear and explain which one is rejected and why. Note that the choice of the two affects the solution -- if we want to address the first problem then log.retention.ms should be used as lower bound on the actual retention time, and if we want to address the second problem then the log.retention.ms should be used as higher bound on the actual retention time.
In both cases, we probably need to figure out a way to determine "active consumer group". Maybe we can compare the time-since-last-commit against a threshold to determine this. In addition, the threshold can be overridden either per-topic or per-groupId. If we go along this route, the rejected solution (per-topic vs. per-groupId) should probably be explained in the KIP. Thanks, Dong On Thu, Oct 13, 2016 at 10:23 AM, Dong Lin <lindon...@gmail.com> wrote: > Hi David, > > Thanks for your explanation. There still seems to be issue with this > solution. Please see my comment inline. > > > On Thu, Oct 13, 2016 at 8:46 AM, 东方甲乙 <254479...@qq.com> wrote: > >> Hi Dong, >> Sorry for the delay, here are the comments: >> 1.I think we should distinguish these two cases: >> (1) group has no member, but has commit offset : In this case we should >> consider its commit offset >> (2) group has no member, no commit offset: Skip this group >> Is it ok? >> >> >> ListGroup API can list the groups, but this API only show the Online >> Group, so we should enhance the listGroup API to list those groups in the >> case (1) >> >> Say some user starts a consumer to consume topic A with > enable.auto.commit = true. Later they change the group name in the config. > Then the proposed solution will never execute consumed log retention for > the topic A, right? I think group name change is pretty common and we > should take care of this issue. One possible solution is to add a config to > specify the maximum time since last offset commit before we consider a > group is inactive. > > >> >> 2. Because every consumer group may appear in different time, say, group >> 1 start to consume in day 1, group 2 start to consume in day 2. If we >> delete the log segment right away, >> group 2 can not consume these message. So we hope the messages can hold >> for a specified time. I think many use-cases will need there configs, if >> there are many consumer groups. >> >> > If we want to take care of group 2, can we simply disable consumed log > retention for the topic and set log retention to 1 day? Can you explain the > benefit of enabling consumed log retention and set consumed log retention > to 1 day? > > Currently the flow graph in the KIP suggests that we delete data iff > (consumed log retention is triggered OR forced log retention is triggered). > And alternative solution is to delete data iff ( (consumed log retention is > disabled OR consumed log retention is triggered) AND forced log retention > is triggered). I would argue that the 2nd scheme is better. Say the > consumed log retention is enabled. The 1st scheme basically interprets > forced log retention as the upper bound of the time the data can stay in > Kafka. The 2nd scheme interprets forced log retention as the lower bound of > the time the data can stay in Kafka, which is more consistent with the > purpose of having this forced log retention (to save disk space). And if we > adopt the 2nd solution, the use-case you suggested can be easily addressed > by setting forced log retention to 1 day and enable consumed log retention. > What do you think? > > > >> >> Thanks, >> David >> >> >> >> >> ------------------ 原始邮件 ------------------ >> 发件人: "Dong Lin";<lindon...@gmail.com>; >> 发送时间: 2016年10月10日(星期一) 下午4:05 >> 收件人: "dev"<dev@kafka.apache.org>; >> >> 主题: Re: [DISCUSS] KIP-68 Add a consumed log retention before log retention >> >> >> >> Hey David, >> >> Thanks for reply. Please see comment inline. >> >> On Mon, Oct 10, 2016 at 12:40 AM, Pengwei (L) <pengwei...@huawei.com> >> wrote: >> >> > Hi Dong >> > Thanks for the questions: >> > >> > 1. Now we don't distinguish inactive or active groups. Because in some >> > case maybe inactive group will become active again, and using the >> previous >> > commit offset. >> > >> > So we will not delete the log segment in the consumer retention if there >> > are some groups consume but not commit, but the log segment can be >> delete by >> > the force retention. >> > >> >> So in the example I provided, the consumed log retention will be >> effectively disabled, right? This seems to be a real problem in operation >> -- we don't want log retention to be un-intentionally disabled simply >> because someone start a tool to consume from that topic. Either this KIP >> should provide a way to handle this, or there should be a way for operator >> to be aware of such case and be able to re-eanble consumed log retention >> for the topic. What do you think? >> >> >> >> > 2. These configs are used to determine the out of date time of the >> > consumed retention, like the parameters of the force retention >> > (log.retention.hours, log.retention.minutes, log.retention.ms). For >> > example, users want the save the log for 3 days, after 3 days, kafka >> will >> > delete the log segments which are >> > >> > consumed by all the consumer group. The log retention thread need these >> > parameters. >> > >> > It makes sense to have configs such as log.retention.ms -- it is used >> to >> make data available for up to a configured amount of time before it is >> deleted. My question is what is the use-case for making log available for >> another e.g. 3 days after it has been consumed by all consumer groups. The >> purpose of this KIP is to allow log to be deleted right as long as all >> interested consumer groups have consumed it. Can you provide a use-case >> for >> keeping log available for longer time after it has been consumed by all >> groups? >> >> >> > >> > Thanks, >> > David >> > >> > >> > > Hey David, >> > > >> > > Thanks for the KIP. Can you help with the following two questions: >> > > >> > > 1) If someone start a consumer (e.g. kafka-console-consumer) to >> consume a >> > > topic for debug/validation purpose, a randome consumer group may be >> > created >> > > and offset may be committed for this consumer group. If no offset >> commit >> > is >> > > made for this consumer group in the future, will this effectively >> > > disable consumed log retention for this topic? In other words, how do >> > this >> > > KIP distinguish active consumer group from inactive ones? >> > > >> > > 2) Why do we need new configs such as log.retention.commitoffset.hou >> rs? >> > Can >> > >we simply delete log segments if consumed log retention is enabled for >> > this >> > > topic and all consumer groups have consumed messages in the log >> segment? >> > > >> > > Thanks, >> > > Dong >> > > >> > > >> > > >> > >On Sat, Oct 8, 2016 at 2:15 AM, Pengwei (L) <pengwei...@huawei.com> >> > wrote: >> > > >> > > > Hi Becket, >> > > > >> > > > Thanks for the feedback: >> > > > 1. We use the simple consumer api to query the commit offset, so we >> > don't >> > > > need to specify the consumer group. >> > > > 2. Every broker using the simple consumer api(OffsetFetchKey) to >> query >> > > > the commit offset in the log retention process. The client can >> commit >> > > > offset or not. >> > > > 3. It does not need to distinguish the follower brokers or leader >> > > > brokers, every brokers can query. >> > > > 4. We don't need to change the protocols, we mainly change the log >> > > > retention process in the log manager. >> > > > >> > > > One question is the query min offset need O(partitions * groups) >> time >> > > > complexity, another alternative is to build an internal topic to >> save >> > every >> > > > partition's min offset, it can reduce to O(1). >> > > > I will update the wiki for more details. >> > > > >> > > > Thanks, >> > > > David >> > > > >> > > > >> > > > > Hi Pengwei, >> > > > > >> > > > > Thanks for the KIP proposal. It is a very useful KIP. At a high >> > level, >> > > > the >> > > > > proposed behavior looks reasonable to me. >> > > > > >> > > > > However, it seems that some of the details are not mentioned in >> the >> > KIP. >> > > > > For example, >> > > > > >> > > > > 1. How will the expected consumer group be specified? Is it >> through >> > a per >> > > > > topic dynamic configuration? >> > > > > 2. How do the brokers detect the consumer offsets? Is it required >> > for a >> > > > > consumer to commit offsets? >> > > > > 3. How do all the replicas know the about the committed offsets? >> > e.g. 1) >> > > > > non-coordinator brokers which do not have the committed offsets, >> 2) >> > > > > follower brokers which do not have consumers directly consuming >> from >> > it. >> > > > > 4. Is there any other changes need to be made (e.g. new >> protocols) in >> > > > > addition to the configuration change? >> > > > > >> > > > > It would be great if you can update the wiki to have more details. >> > > > > >> > > > > Thanks, >> > > > > >> > > > > Jiangjie (Becket) Qin >> > > > > >> > > > > On Wed, Sep 7, 2016 at 2:26 AM, Pengwei (L) < >> pengwei...@huawei.com> >> > > > wrote: >> > > > > >> > > > > > Hi All, >> > > > > > I have made a KIP to enhance the log retention, details as >> > follows: >> > > > > > https://cwiki.apache.org/confluence/display/KAFKA/KIP- >> > > > > > 68+Add+a+consumed+log+retention+before+log+retention >> > > > > > Now start a discuss thread for this KIP , looking forward to >> the >> > > > > > feedback. >> > > > > > >> > > > > > Thanks, >> > > > > > David >> > > > > > >> > > > > > >> > >> > >> > >
