Thanks for the reply, Jun. Currently the byte rate quota does not apply to HeartbeatRequest, JoinGroupRequest/SyncGroupRequest. So the only case those requests are throttled is because the request quota is violated. In that case, the throttle time does not really matter whether we return a full FetchResponse or an empty one. Would it be more consistent if we throttle based on the actual throttle time / channel mute time?
Thanks, Jiangjie (Becket) Qin On Tue, Jan 16, 2018 at 3:45 PM, Jun Rao <j...@confluent.io> wrote: > Hi, Jiangjie, > > You are right that the heartbeat is done in a channel different from the > fetch request. I think it's still useful to return an empty fetch response > when the quota is violated. This way, the throttle time for the heartbeat > request won't be large. I agree that we can just mute the channel for the > fetch request for the throttle time computed based on a full fetch > response. This probably also partially addresses Rajini's #1 concern. > > Thanks, > > Jun > > On Mon, Jan 15, 2018 at 9:27 PM, Becket Qin <becket....@gmail.com> wrote: > > > Hi Rajini, > > > > Thanks for the comments. Pleas see the reply inline. > > > > Hi Jun, > > > > Thinking about the consumer rebalance case a bit more, I am not sure if > we > > need to worry about the delayed rebalance due to quota violation or not. > > The rebalance actually uses a separate channel to coordinator. Therefore > > unless the request quota is hit, the rebalance won't be throttled. Even > if > > request quota is hit, it seems unlikely to be delayed long. So it looks > > that we don't need to unmute the channel earlier than needed. Does this > > address your concern? > > > > Thanks, > > > > Jiangjie (Becket) Qin > > > > On Mon, Jan 15, 2018 at 4:22 AM, Rajini Sivaram <rajinisiva...@gmail.com > > > > wrote: > > > > > Hi Becket, > > > > > > A few questions: > > > > > > 1. KIP says: *Although older client implementations (prior to knowledge > > of > > > this KIP) will immediately send the next request after the broker > > responds > > > without paying attention to the throttle time field, the broker is > > > protected by virtue of muting the channel for time X. i.e., the next > > > request will not be processed until the channel is unmuted. * > > > For fetch requests, the response is sent immediately and the mute time > on > > > the broker based on empty fetch response will often be zero (unlike the > > > throttle time returned to the client based on non-empty response). > Won't > > > that lead to a tight loop of fetch requests from old clients > > (particularly > > > expensive with SSL)? Wouldn't it be better to retain current behaviour > > for > > > old clients? Also, if we change the behaviour for old clients, client > > > metrics that track throttle time will report a lot of throttle > unrelated > > to > > > actual throttle time. > > > > > For consumers, if quota is violated, the throttle time on the broker will > > not be 0. It is just that the throttle time will not be increasing > because > > the consumer will return an empty response in this case. So there should > > not be a tight loop. > > > > > > > 2. KIP says: *The usual idle timeout i.e., connections.max.idle.ms > > > <http://connections.max.idle.ms> will still be honored during the > > throttle > > > time X. This makes sure that the brokers will detect client connection > > > closure in a bounded time.* > > > Wouldn't it be better to bound maximum throttle time to > > > *connections.max.idle.ms > > > <http://connections.max.idle.ms>*? If we mute for a time greater than > > > *connections.max.idle.ms > > > <http://connections.max.idle.ms>* and then close a client connection > > > simply > > > because we had muted it on the broker for a longer throttle time, we > > force > > > a reconnection and read the next request from the new connection > straight > > > away. This feels the same as a bound on the quota of * > > > connections.max.idle.ms > > > <http://connections.max.idle.ms>*, but with added load on the broker > for > > > authenticating another connection (expensive with SSL). > > > > > I think we need to think about the consumer prior to and after this KIP > > separately. > > > > For consumer prior to this KIP, even if we unmute the channel after > > connection.max.idle.ms, it is likely that the consumers have already > > reached request.timeout.ms before that and has reconnected to the > broker. > > So there is no real difference whether we close the throttled channel or > > not. > > > > For consumers after the KIP, because they will honor the throttle time, > > they will back off until throttle time is reached. If that throttle time > is > > longer than connection.max.idle.ms, it seems not a big overhead because > > there will only be one connection re-establishment in quite a few > minutes. > > Compared with such overhead, it seems more important to honor the quota > so > > the broker can survive. > > > > > > > 3. Are we changing the behaviour of network bandwidth quota for > > > Produce/Fetch and retaining current behaviour for request quotas? > > > > > This is going to be applied to all the quota. Including request quotas. > > > > > > > > > > Thanks, > > > > > > Rajini > > > > > > > > > > > > On Wed, Jan 10, 2018 at 10:29 PM, Jun Rao <j...@confluent.io> wrote: > > > > > > > Hi, Jiangjie, > > > > > > > > Thanks for the updated KIP. +1 > > > > > > > > Jun > > > > > > > > On Mon, Jan 8, 2018 at 7:45 PM, Becket Qin <becket....@gmail.com> > > wrote: > > > > > > > > > Thanks for the comments, Jun. > > > > > > > > > > 1. Good point. > > > > > 2. Also makes sense. Usually the connection.max.idle.ms is high > > enough > > > > so > > > > > the throttling is impacted. > > > > > > > > > > I have updated the KIP to reflect the changes. > > > > > > > > > > Thanks, > > > > > > > > > > Jiangjie (Becket) Qin > > > > > > > > > > > > > > > On Mon, Jan 8, 2018 at 6:30 PM, Jun Rao <j...@confluent.io> wrote: > > > > > > > > > > > Hi, Jiangjie, > > > > > > > > > > > > Sorry for the late response. The proposal sounds good overall. A > > > couple > > > > > of > > > > > > minor comments below. > > > > > > > > > > > > 1. For throttling a fetch request, we could potentially just send > > an > > > > > empty > > > > > > response. We can return a throttle time calculated from a full > > > > response, > > > > > > but only mute the channel on the server based on a throttle time > > > > > calculated > > > > > > based on the empty response. This has the benefit that the server > > > will > > > > > mute > > > > > > the channel much shorter, which will prevent the consumer from > > > > > rebalancing > > > > > > when throttled. > > > > > > > > > > > > 2. The wiki says "connections.max.idle.ms should be ignored > during > > > the > > > > > > throttle time X." This has the potential issue that a server may > > not > > > > > detect > > > > > > that a client connection is already gone until after an arbitrary > > > > amount > > > > > of > > > > > > time. Perhaps we could still just close a connection if the > server > > > has > > > > > > muted it for longer than connections.max.idle.ms. This will at > > least > > > > > bound > > > > > > the time for a server to detect closed client connections. > > > > > > > > > > > > Thanks, > > > > > > > > > > > > Jun > > > > > > > > > > > > > > > > > > On Mon, Nov 20, 2017 at 5:30 PM, Becket Qin < > becket....@gmail.com> > > > > > wrote: > > > > > > > > > > > > > Hi, > > > > > > > > > > > > > > We would like to start the voting thread for KIP-219. The KIP > > > > proposes > > > > > to > > > > > > > improve the quota communication between the brokers and > clients, > > > > > > especially > > > > > > > for cases of long throttling time. > > > > > > > > > > > > > > The KIP wiki is following: > > > > > > > https://cwiki.apache.org/confluence/display/KAFKA/KIP- > > > > > > 219+-+Improve+quota+ > > > > > > > communication > > > > > > > > > > > > > > The discussion thread is here: > > > > > > > http://markmail.org/search/?q=kafka+KIP-219#query:kafka% > > > > > > > 20KIP-219+page:1+mid:ooxabguy7nz7l7zy+state:results > > > > > > > > > > > > > > Thanks, > > > > > > > > > > > > > > Jiangjie (Becket) Qin > > > > > > > > > > > > > > > > > > > > > > > > > > > >
