Can you start another thread with [VOTE] in the subject ? -------- Original message --------From: 东方甲乙 <254479...@qq.com> Date: 4/7/18 6:10 AM (GMT-08:00) To: dev <dev@kafka.apache.org> Subject: Re: [DISCUSS] KIP-148: Add a connect timeout for client Hi all,
Do you have further concerns or questions about this KIP? If not, would you please vote for it? Thank you very much! Best, David ------------------ 原始邮件 ------------------ 发件人: "东方甲乙"<254479...@qq.com>; 发送时间: 2017年6月16日(星期五) 晚上10:45 收件人: "dev"<dev@kafka.apache.org>; 主题: Re: [DISCUSS] KIP-148: Add a connect timeout for client Hi Colin, I think the exponential backoff should still apply, thanks for the explanation. thanks, David ------------------ 原始邮件 ------------------ 发件人: "Colin McCabe";<cmcc...@apache.org>; 发送时间: 2017年6月13日(星期二) 凌晨1:43 收件人: "dev"<dev@kafka.apache.org>; 主题: Re: 回复:Re: [DISCUSS] KIP-148: Add a connect timeout for client Just a note: KIP-144 added exponential backoff for broker reconnect attempts, configured via reconnect.backoff.max.ms. cheers, Colin On Sat, Jun 10, 2017, at 08:42, 东方甲乙 wrote: > ------------------ 原始邮件 ------------------ > 发件人: "东方甲乙";<254479...@qq.com>; > 发送时间: 2017年6月4日(星期天) 晚上6:05 > 收件人: "dev"<dev@kafka.apache.org>; > > 主题: Re: [DISCUSS] KIP-148: Add a connect timeout for client > > > > >I guess one obvious question is, how does this interact with retries? > >Does it result in a failure getting delivered to the end user more > >quickly if connecting is impossible the first few times we try? Does > >exponential backoff still apply? > > > Yes, for the retries it will make the end user more quickly to connect. > After the produce request > failed because of timeout, network client close the connection and start > to connect to the leastLoadedNode node. > If the node has no response, we will quickly close the connecting in the > specified timeout and try another node. > > > And for the exponential backoff, do you mean for the TCP's exponential > backoff or the NetworkClient's exponential backoff ? > It seems the NetworkClient has no exponential backoff (the > reconnect.backoff.ms parameter) > > > Thanks > David > > > > > ------------------ 原始邮件 ------------------ > 发件人: "Colin McCabe";<cmcc...@apache.org>; > 发送时间: 2017年5月31日(星期三) 凌晨2:44 > 收件人: "dev"<dev@kafka.apache.org>; > > 主题: Re: [DISCUSS] KIP-148: Add a connect timeout for client > > > > On Mon, May 29, 2017, at 15:46, Guozhang Wang wrote: > > On Wed, May 24, 2017 at 9:59 AM, Colin McCabe <cmcc...@apache.org> wrote: > > > > > On Tue, May 23, 2017, at 19:07, Guozhang Wang wrote: > > > > I think using a single config to cover end-to-end latency with > > > > connecting > > > > and request round-trip may not be best appropriate since 1) some request > > > > may need much more time than others since they are parked (fetch request > > > > with long polling, join group request etc) or throttled, > > > > > > Hmm. My proposal was to implement _both_ end-to-end timeouts and > > > per-call timeouts. In that case, some requests needing much more time > > > than others should not be a concern, since we can simply set a higher > > > per-call timeout on the requests we think will need more time. > > > > > > > and 2) some > > > > requests are prerequisite of others, like group request to discover the > > > > coordinator before the fetch offset request, and implementation wise > > > > these > > > > request send/receive is embedded in latter ones, hence it is not clear > > > > if > > > > the `request.timeout.ms` should cover just a single RPC or more. > > > > > > As far as I know, the request timeout has always covered a single RP If > > > we want to implement a higher level timeout that spans multiple RPCs, we > > > can set the per-call timeouts appropriately. For example: > > > > > > > long deadline = System.currentTimeMillis() + 60000; > > > > callA(callTimeout = deadline - System.currentTimeMillis()) > > > > callB(callTimeout = deadline - System.currentTimeMillis()) > > > > > > > > I may have misunderstand your previous email. Just clarifying: > > > > 1) On the client we already have some configs for controlling end-to-end > > timeout, e.g. "max.block.ms" on producer controls how long "send()" and > > "partitionsFor()" will block for, and inside such API calls multiple > > request round trips may be sent, and for the first request round trip, a > > connecting phase may or may not be included. All of these are be covered > > in > > this "max.block.ms" timeout today. However, as we discussed before not > > all > > request round trips have similar latency expectation, so it is better to > > make a per-request "request.timeout.ms" and the overall "max.block.ms" > > would need to be at least the max of them. > > That makes sense. > > Just to be clear, when you say "per-request timeout" are you talking > about a timeout that can be different for each request? (This doesn't > exist today, but has been proposed.) Or are you talking about > request.timeout.ms, the single timeout that currently applies to all > requests in NetworkClient? > > > > > 2) Now back to the question whether we should make "request.timeout.ms" > > include potential connection phase as well: assume we are going to add > > the > > pre-request "request.timeout.ms" as suggested above, then we may still > > have > > a tight bound on how long connecting should take. For example, let's say > > we > > make "joingroup.request.timeout.ms" (or "fetch.request.timeout.ms" to be > > large since we want really long polling behavior) to be a large value, > > say > > 200 seconds, then if the client is trying to connect to the broker while > > sending the request, and the broker has died, then we may still be > > blocked > > waiting for 30 seconds while I think David's motivation is to fail-fast > > in > > these cases. > > Thanks for the explanation. I think I understand better now. David > wants to be able to have a long timeout for waiting for the server to > process the request, but a shorter timeout for waiting for the > connection to be established. In that case, implementing the additional > timeout makes sense. > > I guess one obvious question is, how does this interact with retries? > Does it result in a failure getting delivered to the end user more > quickly if connecting is impossible the first few times we try? Does > exponential backoff still apply? > > best, > Colin > > > > > > > > > > > > > > So no matter whether we add a `connect.timeout.ms` in addition to ` > > > > request.timeout.ms`, we should consider adding per-request-type timeout > > > > value, and make `request.timeout.ms` a global default; if we add the ` > > > > connect.timeout.ms` the per-request value is only for the round trip, > > > > otherwise it is supposed to include the connecting time. Personally I'd > > > > prefer the first option to add a universal `connect.timeout.ms`, and in > > > > another KIP consider adding per-request-type timeout overrides. > > > > > > Why have a special case for time spent connecting, though? Why would > > > the user care where the time went, as long as the timeout was met? It > > > feels like this is just a hack because we couldn't raise > > > request.timeout.ms to the value that it "should" have been at for the > > > shorter requests. As someone already commented, it's confusing to have > > > all these knobs that we don't really need. > > > > > > > > I think that is exactly what David cares (please correct me if I'm > > wrong): > > for some request I would like to wait long enough for it to be completed, > > like join-group request; while at the same time if it has encountered > > some > > issues while trying to connect to the broker to send the join group > > request, I want to be notified sooner. > > > > > > > > > > > > BTW if the consumer issue is the only cause that we are having a high > > > > default value, I'd suggest we separate the consumer rebalance timeout > > > > and > > > > not piggy-back on the session timeout. Then we can set the default ` > > > > request.timeout.ms` to a smaller value, like 10 secs. This is orthogonal > > > > to > > > > this KIP discussion and we can continue this in a separate thread. > > > > > > +1 > > > > > > cheers, > > > Colin > > > > > > > > > > > > > > > Guozhang > > > > > > > > On Tue, May 23, 2017 at 3:31 PM, Colin McCabe <cmcc...@apache.org> > > > wrote: > > > > > > > > > Another note-- it would be really nice if timeouts were end-to-end, > > > > > rather than being set for particular phases of an RP From a user > > > > > point > > > > > of view, a 30 second timeout should mean that the call either succeeds > > > > > or fails after 30 seconds, regardless of how much time is spent > > > > > looking > > > > > for metadata, connecting to brokers, waiting for brokers, etc. This > > > > > is > > > > > implemented in AdminClient by setting a deadline when the call is > > > > > first > > > > > created and referring to that afterwards. > > > > > > > > > > best, > > > > > Colin > > > > > > > > > > > > > > > On Tue, May 23, 2017, at 13:18, Colin McCabe wrote: > > > > > > In the AdminClient, we allow setting per-call timeouts. The global > > > > > > timeout is just a default. It seems like that is really what we > > > should > > > > > > do in the producer and consumer as well, rather than having a lot of > > > > > > special cases for timeouts in connecting vs. other call states. > > > Then > > > > > > join requests could gave a 5 minute timeout, but other requests > > > > > > could > > > > > > gave a shorter one. Thoughts? > > > > > > > > > > > > Cheers, > > > > > > Colin > > > > > > > > > > > > OnTue, May 23, 2017, at 04:27, Rajini Sivaram wrote: > > > > > > > Guozhang, > > > > > > > > > > > > > > At the moment we don't have a connect timeout. And the behaviour > > > > > > > suggested > > > > > > > in the KIP is useful to address this. > > > > > > > > > > > > > > We do however have a request.timeout.ms. This is the amount of > > > time it > > > > > > > would take to detect a crashed broker if the broker crashed after > > > > > > > a > > > > > > > connection was established. Unfortunately in the consumer, this > > > > > > > was > > > > > > > increased to > 5minutes since JoinRequest can take up to > > > > > > > max.poll.interval.ms, which has a default of 5 minutes. Since the > > > > > > > whole point of this timeout is to detect a crashed broker, 5 > > > minutes is > > > > > > > too > > > > > > > large. > > > > > > > > > > > > > > My suggestion was to use request.timeout.ms to also detect > > > connection > > > > > > > timeouts to a crashed broker - implement the behavior suggested in > > > the > > > > > > > KIP > > > > > > > without adding a new config parameter. As Ismael has said, this > > > will > > > > > need > > > > > > > to fix request.timeout.ms in the consumer. > > > > > > > > > > > > > > > > > > > > > On Mon, May 22, 2017 at 1:23 PM, Simon Souter < > > > > > sim...@cakesolutions.net> > > > > > > > wrote: > > > > > > > > > > > > > > > The following tickets are probably relevant to this KIP: > > > > > > > > > > > > > > > > https://issues.apache.org/jira/browse/KAFKA-3457 > > > > > > > > https://issues.apache.org/jira/browse/KAFKA-1894 > > > > > > > > https://issues.apache.org/jira/browse/KAFKA-3834 > > > > > > > > > > > > > > > > On 22 May 2017 at 16:30, Rajini Sivaram <rajinisiva...@gmail.com > > > > > > > > > wrote: > > > > > > > > > > > > > > > > > Ismael, > > > > > > > > > > > > > > > > > > Yes, agree. My concern was that a connection can be shutdown > > > > > uncleanly at > > > > > > > > > any time. If a client is in the middle of a request, then it > > > times > > > > > out > > > > > > > > > after min(request.timeout.ms, tcp-timeout). If we add another > > > > > config > > > > > > > > > option > > > > > > > > > connect.timeout.ms, then we will sometimes wait for min( > > > > > > > > connect.timeout.ms > > > > > > > > > , > > > > > > > > > tcp-timeout) and sometimes for min(request.timeout.ms, > > > > > tcp-timeout), > > > > > > > > > depending > > > > > > > > > on connection state. One config option feels neater to me. > > > > > > > > > > > > > > > > > > On Mon, May 22, 2017 at 11:21 AM, Ismael Juma < > > > ism...@juma.me.uk> > > > > > wrote: > > > > > > > > > > > > > > > > > > > Rajini, > > > > > > > > > > > > > > > > > > > > For this to have the desired effect, we'd probably need to > > > lower > > > > > the > > > > > > > > > > default request.timeout.ms for the consumer and fix the > > > > > underlying > > > > > > > > > reason > > > > > > > > > > why it is a little over 5 minutes at the moment. > > > > > > > > > > > > > > > > > > > > Ismael > > > > > > > > > > > > > > > > > > > > On Mon, May 22, 2017 at 4:15 PM, Rajini Sivaram < > > > > > > > > rajinisiva...@gmail.com > > > > > > > > > > > > > > > > > > > > wrote: > > > > > > > > > > > > > > > > > > > > > Hi David, > > > > > > > > > > > > > > > > > > > > > > Sorry, what I meant was: Can you reuse the existing > > > > > configuration > > > > > > > > > option > > > > > > > > > > > request.timeout,ms , instead of adding a new config and > > > add the > > > > > > > > > behaviour > > > > > > > > > > > that you have proposed in the KIP for the connection phase > > > > > using this > > > > > > > > > > > timeout? I think the timeout for connection is useful. I > > > > > > > > > > > am > > > > > not sure > > > > > > > > we > > > > > > > > > > > need another configuration option to implement it. > > > > > > > > > > > > > > > > > > > > > > Regards, > > > > > > > > > > > > > > > > > > > > > > Rajini > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > On Mon, May 22, 2017 at 11:06 AM, 东方甲乙 <254479...@qq.com> > > > > > wrote: > > > > > > > > > > > > > > > > > > > > > > > Hi Rajini. > > > > > > > > > > > > > > > > > > > > > > > > When kafka node' machine is shutdown or network is > > > closed, > > > > > the > > > > > > > > > > connecting > > > > > > > > > > > > phase could not use the request.timeout.ms, because the > > > > > client > > > > > > > > > haven't > > > > > > > > > > > > send a req yet. And no response for the nio, the > > > selector > > > > > will > > > > > > > > not > > > > > > > > > > > close > > > > > > > > > > > > the connect, so it will not choose other good node to > > > get the > > > > > > > > > metadata. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Thanks > > > > > > > > > > > > David > > > > > > > > > > > > > > > > > > > > > > > > ------------------ 原始邮件 ------------------ > > > > > > > > > > > > *发件人:* "Rajini Sivaram" <rajinisiva...@gmail.com>; > > > > > > > > > > > > *发送时间:* 2017年5月22日(星期一) 20:17 > > > > > > > > > > > > *收件人:* "dev" <dev@kafka.apache.org>; > > > > > > > > > > > > *主题:* Re: [DISCU