Hi, Rajini, Thanks for the proposal.
The benefit of using the request processing time over the request rate is exactly what people have said. I will just expand that a bit. Consider the following case. The producer sends a produce request with a 10MB message but compressed to 100KB with gzip. The decompression of the message on the broker could take 10-15 seconds, during which time, a request handler thread is completely blocked. In this case, neither the byte-in quota nor the request rate quota may be effective in protecting the broker. Consider another case. A consumer group starts with 10 instances and later on switches to 20 instances. The request rate will likely double, but the actually load on the broker may not double since each fetch request only contains half of the partitions. Request rate quota may not be easy to configure in this case. What we really want is to be able to prevent a client from using too much of the server side resources. In this particular KIP, this resource is the capacity of the request handler threads. I agree that it may not be intuitive for the users to determine how to set the right limit. However, this is not completely new and has been done in the container world already. For example, Linux cgroup (https://access.redhat.com/ documentation/en-US/Red_Hat_Enterprise_Linux/6/html/ Resource_Management_Guide/sec-cpu.html) has the concept of cpu.cfs_quota_us, which specifies the total amount of time in microseconds for which all tasks in a cgroup can run during a one second period. We can potentially model the request handler threads in a similar way. For example, each request handler thread can be 1 request handler unit and the admin can configure a limit on how many units (say 0.01) a client can have. Regarding not throttling the internal broker to broker requests. We could do that. Alternatively, we could just let the admin configure a high limit for the kafka user (it may not be able to do that easily based on clientId though). Ideally we want to be able to protect the utilization of the network thread pool too. The difficult is mostly what Rajini said: (1) The mechanism for throttling the requests is through Purgatory and we will have to think through how to integrate that into the network layer. (2) In the network layer, currently we know the user, but not the clientId of the request. So, it's a bit tricky to throttle based on clientId there. Plus, the byteOut quota can already protect the network thread utilization for fetch requests. So, if we can't figure out this part right now, just focusing on the request handling threads for this KIP is still a useful feature. Thanks, Jun On Tue, Feb 21, 2017 at 4:27 AM, Rajini Sivaram <rajinisiva...@gmail.com> wrote: > Thank you all for the feedback. > > Jay: I have removed exemption for consumer heartbeat etc. Agree that > protecting the cluster is more important than protecting individual apps. > Have retained the exemption for StopReplicat/LeaderAndIsr etc, these are > throttled only if authorization fails (so can't be used for DoS attacks in > a secure cluster, but allows inter-broker requests to complete without > delays). > > I will wait another day to see if these is any objection to quotas based on > request processing time (as opposed to request rate) and if there are no > objections, I will revert to the original proposal with some changes. > > The original proposal was only including the time used by the request > handler threads (that made calculation easy). I think the suggestion is to > include the time spent in the network threads as well since that may be > significant. As Jay pointed out, it is more complicated to calculate the > total available CPU time and convert to a ratio when there *m* I/O threads > and *n* network threads. ThreadMXBean#getThreadCPUTime() may give us what > we want, but it can be very expensive on some platforms. As Becket and > Guozhang have pointed out, we do have several time measurements already for > generating metrics that we could use, though we might want to switch to > nanoTime() instead of currentTimeMillis() since some of the values for > small requests may be < 1ms. But rather than add up the time spent in I/O > thread and network thread, wouldn't it be better to convert the time spent > on each thread into a separate ratio? UserA has a request quota of 5%. Can > we take that to mean that UserA can use 5% of the time on network threads > and 5% of the time on I/O threads? If either is exceeded, the response is > throttled - it would mean maintaining two sets of metrics for the two > durations, but would result in more meaningful ratios. We could define two > quota limits (UserA has 5% of request threads and 10% of network threads), > but that seems unnecessary and harder to explain to users. > > Back to why and how quotas are applied to network thread utilization: > a) In the case of fetch, the time spent in the network thread may be > significant and I can see the need to include this. Are there other > requests where the network thread utilization is significant? In the case > of fetch, request handler thread utilization would throttle clients with > high request rate, low data volume and fetch byte rate quota will throttle > clients with high data volume. Network thread utilization is perhaps > proportional to the data volume. I am wondering if we even need to throttle > based on network thread utilization or whether the data volume quota covers > this case. > > b) At the moment, we record and check for quota violation at the same time. > If a quota is violated, the response is delayed. Using Jay'e example of > disk reads for fetches happening in the network thread, We can't record and > delay a response after the disk reads. We could record the time spent on > the network thread when the response is complete and introduce a delay for > handling a subsequent request (separate out recording and quota violation > handling in the case of network thread overload). Does that make sense? > > > Regards, > > Rajini > > > On Tue, Feb 21, 2017 at 2:58 AM, Becket Qin <becket....@gmail.com> wrote: > > > Hey Jay, > > > > Yeah, I agree that enforcing the CPU time is a little tricky. I am > thinking > > that maybe we can use the existing request statistics. They are already > > very detailed so we can probably see the approximate CPU time from it, > e.g. > > something like (total_time - request/response_queue_time - remote_time). > > > > I agree with Guozhang that when a user is throttled it is likely that we > > need to see if anything has went wrong first, and if the users are well > > behaving and just need more resources, we will have to bump up the quota > > for them. It is true that pre-allocating CPU time quota precisely for the > > users is difficult. So in practice it would probably be more like first > set > > a relative high protective CPU time quota for everyone and increase that > > for some individual clients on demand. > > > > Thanks, > > > > Jiangjie (Becket) Qin > > > > > > On Mon, Feb 20, 2017 at 5:48 PM, Guozhang Wang <wangg...@gmail.com> > wrote: > > > > > This is a great proposal, glad to see it happening. > > > > > > I am inclined to the CPU throttling, or more specifically processing > time > > > ratio instead of the request rate throttling as well. Becket has very > > well > > > summed my rationales above, and one thing to add here is that the > former > > > has a good support for both "protecting against rogue clients" as well > as > > > "utilizing a cluster for multi-tenancy usage": when thinking about how > to > > > explain this to the end users, I find it actually more natural than the > > > request rate since as mentioned above, different requests will have > quite > > > different "cost", and Kafka today already have various request types > > > (produce, fetch, admin, metadata, etc), because of that the request > rate > > > throttling may not be as effective unless it is set very > conservatively. > > > > > > Regarding to user reactions when they are throttled, I think it may > > differ > > > case-by-case, and need to be discovered / guided by looking at relative > > > metrics. So in other words users would not expect to get additional > > > information by simply being told "hey, you are throttled", which is all > > > what throttling does; they need to take a follow-up step and see "hmm, > > I'm > > > throttled probably because of ..", which is by looking at other metric > > > values: e.g. whether I'm bombarding the brokers with metadata request, > > > which are usually cheap to handle but I'm sending thousands per second; > > or > > > is it because I'm catching up and hence sending very heavy fetching > > request > > > with large min.bytes, etc. > > > > > > Regarding to the implementation, as once discussed with Jun, this seems > > not > > > very difficult since today we are already collecting the "thread pool > > > utilization" metrics, which is a single percentage "aggregateIdleMeter" > > > value; but we are already effectively aggregating it for each requests > in > > > KafkaRequestHandler, and we can just extend it by recording the source > > > client id when handling them and aggregating by clientId as well as the > > > total aggregate. > > > > > > > > > Guozhang > > > > > > > > > > > > > > > On Mon, Feb 20, 2017 at 4:27 PM, Jay Kreps <j...@confluent.io> wrote: > > > > > > > Hey Becket/Rajini, > > > > > > > > When I thought about it more deeply I came around to the "percent of > > > > processing time" metric too. It seems a lot closer to the thing we > > > actually > > > > care about and need to protect. I also think this would be a very > > useful > > > > metric even in the absence of throttling just to debug whose using > > > > capacity. > > > > > > > > Two problems to consider: > > > > > > > > 1. I agree that for the user it is understandable what lead to > their > > > > being throttled, but it is a bit hard to figure out the safe range > > for > > > > them. i.e. if I have a new app that will send 200 messages/sec I > can > > > > probably reason that I'll be under the throttling limit of 300 > > > req/sec. > > > > However if I need to be under a 10% CPU resources limit it may be > a > > > bit > > > > harder for me to know a priori if i will or won't. > > > > 2. Calculating the available CPU time is a bit difficult since > there > > > are > > > > actually two thread pools--the I/O threads and the network > threads. > > I > > > > think > > > > it might be workable to count just the I/O thread time as in the > > > > proposal, > > > > but the network thread work is actually non-trivial (e.g. all the > > disk > > > > reads for fetches happen in that thread). If you count both the > > > network > > > > and > > > > I/O threads it can skew things a bit. E.g. say you have 50 network > > > > threads, > > > > 10 I/O threads, and 8 cores, what is the available cpu time > > available > > > > in a > > > > second? I suppose this is a problem whenever you have a bottleneck > > > > between > > > > I/O and network threads or if you end up significantly > > > over-provisioning > > > > one pool (both of which are hard to avoid). > > > > > > > > An alternative for CPU throttling would be to use this api: > > > > http://docs.oracle.com/javase/1.5.0/docs/api/java/lang/ > > > > management/ThreadMXBean.html#getThreadCpuTime(long) > > > > > > > > That would let you track actual CPU usage across the network, I/O > > > threads, > > > > and purgatory threads and look at it as a percentage of total cores. > I > > > > think this fixes many problems in the reliability of the metric. It's > > > > meaning is slightly different as it is just CPU (you don't get > charged > > > for > > > > time blocking on I/O) but that may be okay because we already have a > > > > throttle on I/O. The downside is I think it is possible this api can > be > > > > disabled or isn't always available and it may also be expensive (also > > > I've > > > > never used it so not sure if it really works the way i think). > > > > > > > > -Jay > > > > > > > > On Mon, Feb 20, 2017 at 3:17 PM, Becket Qin <becket....@gmail.com> > > > wrote: > > > > > > > > > If the purpose of the KIP is only to protect the cluster from being > > > > > overwhelmed by crazy clients and is not intended to address > resource > > > > > allocation problem among the clients, I am wondering if using > request > > > > > handling time quota (CPU time quota) is a better option. Here are > the > > > > > reasons: > > > > > > > > > > 1. request handling time quota has better protection. Say we have > > > request > > > > > rate quota and set that to some value like 100 requests/sec, it is > > > > possible > > > > > that some of the requests are very expensive actually take a lot of > > > time > > > > to > > > > > handle. In that case a few clients may still occupy a lot of CPU > time > > > > even > > > > > the request rate is low. Arguably we can carefully set request rate > > > quota > > > > > for each request and client id combination, but it could still be > > > tricky > > > > to > > > > > get it right for everyone. > > > > > > > > > > If we use the request time handling quota, we can simply say no > > clients > > > > can > > > > > take up to more than 30% of the total request handling capacity > > > (measured > > > > > by time), regardless of the difference among different requests or > > what > > > > is > > > > > the client doing. In this case maybe we can quota all the requests > if > > > we > > > > > want to. > > > > > > > > > > 2. The main benefit of using request rate limit is that it seems > more > > > > > intuitive. It is true that it is probably easier to explain to the > > user > > > > > what does that mean. However, in practice it looks the impact of > > > request > > > > > rate quota is not more quantifiable than the request handling time > > > quota. > > > > > Unlike the byte rate quota, it is still difficult to give a number > > > about > > > > > impact of throughput or latency when a request rate quota is hit. > So > > it > > > > is > > > > > not better than the request handling time quota. In fact I feel it > is > > > > > clearer to tell user that "you are limited because you have taken > 30% > > > of > > > > > the CPU time on the broker" than otherwise something like "your > > request > > > > > rate quota on metadata request has reached". > > > > > > > > > > Thanks, > > > > > > > > > > Jiangjie (Becket) Qin > > > > > > > > > > > > > > > On Mon, Feb 20, 2017 at 2:23 PM, Jay Kreps <j...@confluent.io> > wrote: > > > > > > > > > > > I think this proposal makes a lot of sense (especially now that > it > > is > > > > > > oriented around request rate) and fills the biggest remaining gap > > in > > > > the > > > > > > multi-tenancy story. > > > > > > > > > > > > I think for intra-cluster communication (StopReplica, etc) we > could > > > > avoid > > > > > > throttling entirely. You can secure or otherwise lock-down the > > > cluster > > > > > > communication to avoid any unauthorized external party from > trying > > to > > > > > > initiate these requests. As a result we are as likely to cause > > > problems > > > > > as > > > > > > solve them by throttling these, right? > > > > > > > > > > > > I'm not so sure that we should exempt the consumer requests such > as > > > > > > heartbeat. It's true that if we throttle an app's heartbeat > > requests > > > it > > > > > may > > > > > > cause it to fall out of its consumer group. However if we don't > > > > throttle > > > > > it > > > > > > it may DDOS the cluster if the heartbeat interval is set > > incorrectly > > > or > > > > > if > > > > > > some client in some language has a bug. I think the policy with > > this > > > > kind > > > > > > of throttling is to protect the cluster above any individual app, > > > > right? > > > > > I > > > > > > think in general this should be okay since for most deployments > > this > > > > > > setting is meant as more of a safety valve---that is rather than > > set > > > > > > something very close to what you expect to need (say 2 req/sec or > > > > > whatever) > > > > > > you would have something quite high (like 100 req/sec) with this > > > meant > > > > to > > > > > > prevent a client gone crazy. I think when used this way allowing > > > those > > > > to > > > > > > be throttled would actually provide meaningful protection. > > > > > > > > > > > > -Jay > > > > > > > > > > > > > > > > > > > > > > > > On Fri, Feb 17, 2017 at 9:05 AM, Rajini Sivaram < > > > > rajinisiva...@gmail.com > > > > > > > > > > > > wrote: > > > > > > > > > > > > > Hi all, > > > > > > > > > > > > > > I have just created KIP-124 to introduce request rate quotas to > > > > Kafka: > > > > > > > > > > > > > > https://cwiki.apache.org/confluence/display/KAFKA/KIP- > > > > > > > 124+-+Request+rate+quotas > > > > > > > > > > > > > > The proposal is for a simple percentage request handling time > > quota > > > > > that > > > > > > > can be allocated to *<client-id>*, *<user>* or *<user, > > client-id>*. > > > > > There > > > > > > > are a few other suggestions also under "Rejected alternatives". > > > > > Feedback > > > > > > > and suggestions are welcome. > > > > > > > > > > > > > > Thank you... > > > > > > > > > > > > > > Regards, > > > > > > > > > > > > > > Rajini > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > -- > > > -- Guozhang > > > > > >
