Hi, Joe, A couple of comments.
1. When creating a new topic, our replica assignment algorithm tries to achieve a few things: (a) all replicas are spread evenly across brokers; (b) the preferred replica (first replica in the assigned replica list) of all partitions are spread evenly across brokers; (c) the non-preferred replicas are spread out in such a way that if we lose a broker, the load on the failed broker is spread evenly among the remaining brokers. For example, if you look at the following replica assignment on brokers b1, b2, and b3 (with replication factor 2). Broker b1 will be the leader for partition p0 and p3. Broker b2 will be the leader for partition p1 and p4. Broker b3 will be the leader for partition p2 and p5. If b1 is gone, b2 will take over as the leader for p0 and b3 will take over as the leader for p3. This strategy makes sure that the load is even in the normal case as well as the failure case. b1 b2 b3 p0 p1 p2 p2 p0 p1 p3 p4 p5 p4 p5 p3 The current reassignment strategy actually maintains properties (a), (b) and (c) after the reassignment completes. The new algorithm takes the last few replicas from an overloaded broker and moves them to an underloaded broker. It does reduce the data movement compared with the current algorithm. It also maintains property (a). However, it doesn't seem to explicitly maintain properties (b) and (c). Data movement is a one-time cost. Maintaining balance after the data movement has long term benefit. So, it will be useful to try to maintain these properties even perhaps at the expense of a bit more data movement. Also, I think the new algorithm needs to make sure that we don't move the same replica to a new broker more than once. 2. I am not sure that we need to add a new --rebalance option. All we are changing is the assignment strategy. If that's a better strategy than before, there is no reason for anyone to use the old strategy. So, the new strategy should just be used in the --generate mode. Thanks, Jun On Wed, Mar 11, 2015 at 12:12 PM, Joe Stein <joe.st...@stealth.ly> wrote: > Sorry for not catching up on this thread earlier, I wanted to-do this > before the KIP got its updates so we could discuss if need be and not waste > more time re-writing/working things that folks have issues with or such. I > captured all the comments so far here with responses. > > << So fair assignment by count (taking into account the current partition > count of each broker) is very good. However, it's worth noting that all > partitions are not created equal. We have actually been performing more > rebalance work based on the partition size on disk, as given equal > retention of all topics, the size on disk is a better indicator of the > amount of traffic a partition gets, both in terms of storage and network > traffic. Overall, this seems to be a better balance. > > Agreed though this is out of scope (imho) for what the motivations for the > KIP were. The motivations section is blank (that is on me) but honestly it > is because we did all the development, went back and forth with Neha on the > testing and then had to back it all into the KIP process... Its a > time/resource/scheduling and hope to update this soon on the KIP ... all of > this is in the JIRA and code patch so its not like it is not there just not > in the place maybe were folks are looking since we changed where folks > should look. > > Initial cut at "Motivations": the --generate is not used by a lot of folks > because they don't trust it. Issues such as giving different results > sometimes when you run it. Also other feedback from the community that it > does not account for specific uses cases like "adding new brokers" and > "removing brokers" (which is where that patch started > https://issues.apache.org/jira/browse/KAFKA-1678 but then we changed it > after review into just --rebalance > https://issues.apache.org/jira/browse/KAFKA-1792). The use case for add > and > remove brokers is one that happens in AWS and auto scailing. There are > other reasons for this too of course. The goal originally was to make what > folks are already coding today (with the output of " available in the > project for the community. Based on the discussion in the JIRA with Neha we > all agreed that making it be a faire rebalance would fulfill both uses > cases. > > << In addition to this, I think there is very much a need to have Kafka be > rack-aware. That is, to be able to assure that for a given cluster, you > never assign all replicas for a given partition in the same rack. This > would allow us to guard against maintenances or power failures that affect > a full rack of systems (or a given switch). > > Agreed, this though I think is out of scope for this change and something > we can also do in the future. There is more that we have to figure out for > rack aware specifically answering "how do we know what rack the broker is > on". I really really (really) worry that we keep trying to put too much > into a single change the discussions go into rabbit holes and good > important features (that are community driven) that could get out there > will get bogged down with different uses cases and scope creep. So, I think > rack awareness is its own KIP that has two parts... setting broker rack and > rebalancing for that. That features doesn't invalidate the need for > --rebalance but can be built on top of it. > > << I think it would make sense to implement the reassignment logic as a > pluggable component. That way it would be easy to select a scheme when > performing a reassignment (count, size, rack aware). Configuring a default > scheme for a cluster would allow for the brokers to create new topics and > partitions in compliance with the requested policy. > > I don't agree with this because right now you get back "the current state > of the partitions" so you can (today) write whatever logic you want (with > the information that is there). With --rebalance you also get that back so > moving forward. Moving forward we can maybe expose more information so that > folks can write different logic they want > (like partition number, location (label string for rack), size, throughput > average, etc, etc, etc... but again... that to me is a different > KIP entirely from the motivations of this one. If eventually we want to > make it plugable then we should have a KIP and discussion around it I just > don't see how it relates to the original motivations of the change. > > << Is it possible to describe the proposed partition reassignment algorithm > in more detail on the KIP? In fact, it would be really easy to understand > if we had some concrete examples comparing partition assignment with the > old algorithm and the new. > > sure, it is in the JIRA linked to the KIP too though > https://issues.apache.org/jira/browse/KAFKA-1792 and documented in > comments > in the patch also as requested. Let me know if this is what you are looking > for and we can simply update the KIP with this information or give more > detail specifically what you think might be missing please. > > << Would we want to > support some kind of automated reassignment of existing partitions > (personally - no. I want to trigger that manually because it is a very disk > and network intensive process)? > > You can automate the reassignment with a line of code that takes the > response and calls --execute if folks want that... I don't think we should > ever link these (or at least not yet) because of the reasons you say. I > think as long as we have a way > > ******** > > If there is anything else I missed please let me know so I can make sure > that the detail gets update so we minimize the back and forth both in > efforts and elapsed time. This was always supposed to be a very small fix > for something that pains A LOT of people and I want to make sure that we > aren't running scope creep on the change but are making sure that folks > understand the motivation behind a new feature. > > ~ Joe Stein > - - - - - - - - - - - - - - - - - > > http://www.stealth.ly > - - - - - - - - - - - - - - - - - > > On Sun, Mar 8, 2015 at 1:21 PM, Joe Stein <joe.st...@stealth.ly> wrote: > > > Jay, > > > > That makes sense. I think what folks are bringing up all sounds great > but > > I feel can/should be done afterwards as further improvements as the scope > > for this change has a very specific focus to resolve problems folks have > > today with --generate (with a patch tested and ready to go ). I should be > > able to update the KIP this week and followup. > > > > ~ Joestein > > On Mar 8, 2015 12:54 PM, "Jay Kreps" <jay.kr...@gmail.com> wrote: > > > >> Hey Joe, > >> > >> This still seems pretty incomplete. It still has most the sections just > >> containing the default text you are supposed to replace. It is really > hard > >> to understand what is being proposed and why and how much of the problem > >> we > >> are addressing. For example the motivation section just says > >> "operational". > >> > >> I'd really like us to do a good job of this. I actually think putting > the > >> time in to convey context really matters. For example I think (but can't > >> really know) that what you are proposing is just a simple fix to the > JSON > >> output of the command line tool. But you can see that on the thread it > is > >> quickly going to spiral into automatic balancing, rack awareness, data > >> movement throttling, etc. > >> > >> Just by giving people a fairly clear description of the change and how > it > >> fits into other efforts that could happen in the area really helps keep > >> things focused on what you want. > >> > >> -Jay > >> > >> > >> On Wed, Jan 21, 2015 at 10:11 PM, Joe Stein <joe.st...@stealth.ly> > wrote: > >> > >> > Posted a KIP for --re-balance for partition assignment in reassignment > >> > tool. > >> > > >> > > >> > > >> > https://cwiki.apache.org/confluence/display/KAFKA/KIP-6+-+New+reassignment+partition+logic+for+re-balancing > >> > > >> > JIRA https://issues.apache.org/jira/browse/KAFKA-1792 > >> > > >> > While going through the KIP I thought of one thing from the JIRA that > we > >> > should change. We should preserve --generate to be existing > >> functionality > >> > for the next release it is in. If folks want to use --re-balance then > >> > great, it just won't break any upgrade paths, yet. > >> > > >> > /******************************************* > >> > Joe Stein > >> > Founder, Principal Consultant > >> > Big Data Open Source Security LLC > >> > http://www.stealth.ly > >> > Twitter: @allthingshadoop <http://www.twitter.com/allthingshadoop> > >> > ********************************************/ > >> > > >> > > >
