"...I think GC implementation is orthogonal to my proposal. So if you can deal with such problems, I would gladly apply it as well..." I think it's definitely worth trying. The "mark-and-sweep" pattern with an eventual consistent state may give GC a good premise to make decisions. If GC is instantiated as a singleton action, through Akka, it can achieve a highly fault tolerant service, allowing for self-healing with no single point of failure.
If an action changes state every "2ms" as in the example above, and the GC gets that change after "100ms", this could work fine; a timeout of "10mins" may give GC a decent time window to make decisions. The "mark-and-sweep" then allows containers ( ContainerProxy ) to move that container in the "survival" space - to use GC G1 language - in case the container processed an activation in between "mark" and "sweep" commands. If a container "survived", then GC will prolong its TTL with another 10mins, and so forth. At the same time GC doesn't need to know about *all* state changes (busy->warm->etc), especially if they happen so frequently. This generates unnecessary noise in the network. GC could use a sampling strategy instead, or check for "last activation" *only* when the allocated memory in the cluster is over a threshold; there are a number of strategies that could ensure GC runs with "minimum enough info" and "as infrequent as needed" to make the right decisions. "...If we "must" guarantee the execution of users, we need to have a bigger cluster than the sum of users' limit.... I agree. For the cases when an operator can't run such a large cluster to accommodate all namespaces regardless of their traffic, that's when a smart GC becomes critical to achieve cost efficiency. At least in the deployments I'm involved in, the cluster aims to shut-down most of its VMs when there's no traffic, and scale back up as quickly as possible as traffic increases. We run clusters in public clouds; we can't afford to keep thousands of VMs running if nobody uses them. When cluster scales up, there's a time window when there could be a congestion of resources; so, as with network congestion, activations should suffer for a little while, but at least they all get a fair chance to execute. They execute slower, but at least they execute; but if containers are left to destroy themselves, then it's hard to predict cluster behavior when congested, and it's hard to guarantee an SLA. This is where I was coming from. I'm sure we'll find a way to accommodate multiple deployment patterns. I'm describing my setup with the hope that the new architecture will allow a configuration mechanism or SPIs for such key areas. On Thu, Apr 11, 2019 at 8:15 PM Dominic Kim <[email protected]> wrote: > Well, that is not the tradeoff only resides in my proposal. > If we "must" guarantee the execution of users, we need to have a bigger > cluster than the sum of users' limit. > Even though we use current implementation, if the sum of concurrent limit > exceeds the system resources, we cannot completely guarantee all executions > under a burst. > (We cannot guarantee 200 concurrent execution with 100 containers.) > > The reason why I used a kind of self-GC is, it's not easy to track the > resource status in real time. > It is the same with the reason why I take the pull-based model. > > For example, if we control the container deletion in a central way, we > should track all container status such as how many containers are running > for each action, which of them are idle, where they are running, and so on. > But the status of resources changes blazingly fast. > Below is one example. The execution is over within 2 ms. > { > "namespace": "style95", > "name": "hello-world", > "version": "0.0.1", > "subject": "style95", > "activationId": "ba2cc561fc8e4272acc561fc8ea27210", > "start": 1554967214351, > "end": 1554967214353, > "duration": 2, > "response": { > "status": "success", > "statusCode": 0, > . > . > . > } > > It means the container status(busy, warm) can change every 2 ms. > > Also, if you are not thinking of one central component which decides to > delete containers,(it can be a SPOF) there will be multiple components > which are making the decision. > When one of them makes a decision, it should consider decisions made(and > will be made) by the others at the same time. > So we need to track down all container status, consider all decision made > by multiple components and finally make the optimal decision to delete > containers within 2 ms. > I think this is not viable. > > Your idea sounds great, it could be a great enhancement. > And I think GC implementation is orthogonal to my proposal. > So if you can deal with such problems, I would gladly apply it as well. > > > Best regards > Dominic > > > > > > > 2019년 4월 11일 (목) 오후 3:15, Dascalita Dragos <[email protected]>님이 작성: > > > Thanks Dominic for the details. > > > > It seems like an operator has to choose between “do I hurt > performance(low > > timeout) or do I hurt the SLA” ? > > > > If this is the trade off , isn’t this a hard choice to make ? So I’m > > wondering whether some alternative designs could be used for this > problem. > > > > The key decision here is: should OW be given a cluster wide power to view > > and control the resources or not. IIUC the current proposal doesn’t > support > > this? I’m not saying the proposed model is not good; I’d just feel more > > comfortable if OW would allow more options instead of one, in the same > way > > the JVM allows multiple GC implementations. In the proposed model the GC > > would offload the decision to each container, while other implementations > > may do it differently. For instance, I’d implement something dynamic > that > > adapts the timeout to the load, and maybe try some predictive ML > algorithms > > to manage resources - if a model suggests that out of 3 actions that > could > > be removed, 1 has a higher probability to be invoked again, wouldn’t it > be > > more efficient to remove one of the other 2 ? Such a logic can only be > > achieved through an entity with a cluster wide view, as actions don’t > know > > about each other, to negotiate a dynamic timeout. > > > > - dragos > > > > On Wed, Apr 10, 2019 at 3:46 AM Dominic Kim <[email protected]> wrote: > > > > > Dear Dascalita > > > > > > That depends on the timeout configuration. > > > For example, if you need something similar to the one in the current > code > > > base, you can just configure the timeout to a small enough value, such > as > > > 50ms. > > > > > > The idea behind the longer timeout is, it shows better performance when > > > there are highly likely subsequent requests. > > > For example, it takes about 100ms ~ 1s to create a new coldstart > > container. > > > If the action execution takes 10ms, it should wait 10 to 100 times more > > for > > > a new container. > > > In this case, it is reasonable to wait for the previous execution and > > reuse > > > the existing container rather than creating a new container. > > > So 100ms ~ 1s could be good options for the timeout value. > > > (Under heavy loads, I even observed it took 2s ~ 5s to create a > coldstart > > > container.) > > > And this implies some changes in the notion of resources. > > > > > > In the cluster, there would be a different kind of requests. > > > There would be both batch and real-time invocation. > > > So I think this is a tradeoff. > > > Longer timeout will increase the reuse rate of containers but idle > > > containers will possess resources longer. > > > > > > And even in the current implementation, subsequent invocation should > wait > > > for some time to remove existing(warmed containers) and create a new > cold > > > start container. > > > As I said, it could take up to few seconds under heavy loads. > > > With reasonable timeout value, there would be no big performance > > difference > > > in the above situation. > > > (Actually, I expect new scheduler would outperform even with 5~10s > > timeout > > > value as it will evenly distribute docker operation. > > > In the current implementation, all execution is sent to the home > invoker > > > first and it could make the situation worse in edge cases. > > > I hope I can share performance comparison results as I am doing > > > benchmarking.) > > > > > > Also, I think the above case is an edge case that someone is consuming > > most > > > of the cluster resources and executing two different batch invocation > > > alternatively. > > > Anyway, we can support such an edge case with some shutdown period. > > > This can be controversial, but I believe this is a viable option. > > > > > > > > > If you said that in the view of OpenWhisk operator, I think you meant > > there > > > are more than 1 heavy users. > > > Let's say, one user has 60 containers limit and the other has 80 > > containers > > > limit. > > > Then can we guarantee both users' execution without any issue in > current > > > implementation? > > > If their invocation requests come together, one or both of their > > invocation > > > will be heavily delayed. > > > > > > So I think when we(operators) notice there are such heavy users, we > > should > > > scale out our clusters to guarantee their invocation or we should > reduce > > > their resource limit. > > > This is also a tradeoff. If we must guarantee their invocation, we at > > least > > > need a bigger cluster than the sum of their throttling limit. > > > If we have weak SLA, we can support both users with smaller cluster > > though > > > their invocation can be delayed a bit. > > > > > > > > > In short, if you prefer the current behavior you can still have a > similar > > > effect by configuring the timeout as 50ms. > > > (So containers will only wait for 50ms, though it may induce some > > > performance degradation in other cases.) > > > > > > Best regards > > > Dominic > > > > > > > > > 2019년 4월 10일 (수) 오전 1:36, Dascalita Dragos <[email protected]>님이 작성: > > > > > > > "...When there is no more activation message, ContainerProxy will be > > wait > > > > for the given time(configurable) and just stop...." > > > > > > > > How does the system allocate and de-allocate resources when it's > > > congested > > > > ? > > > > I'm thinking at the use case where the system receives a batch of > > > > activations that require 60% of all cluster resources. Once those > > > > activations finish, a different batch of activations are received, > and > > > this > > > > time the new batch requires new actions to be cold-started; these new > > > > activations require a total of 80% of the overall cluster resources. > > > Unless > > > > the previous actions are removed, the cluster is over-allocated. In > the > > > > current model would the cluster process 1/2 of the new activations > b/c > > it > > > > needs to wait for the previous actions to stop by themselves ? > > > > > > > > On Sun, Apr 7, 2019 at 7:34 PM Dominic Kim <[email protected]> > > wrote: > > > > > > > > > Hi Mingyu > > > > > > > > > > Thank you for the good questions. > > > > > > > > > > Before answering to your question, I will share the Lease in ETCD > > > first. > > > > > ETCD has a data model which is disappear after given time if there > is > > > no > > > > > relevant keepalive on it, the Lease. > > > > > > > > > > So once you grant a new lease, you can put it with data in each > > > operation > > > > > such as put, putTxn(transaction), etc. > > > > > If there is no keep-alive for the given(configurable) time, > inserted > > > data > > > > > will be gone. > > > > > > > > > > In my proposal, most of data in ETCD rely on a lease. > > > > > For example, each scheduler stores their endpoint information(for > > queue > > > > > creation) with a lease. Each queue stores their information(for > > > > activation) > > > > > in ETCD with a lease. > > > > > (It is overhead to do keep-alive in each memory queue separately, I > > > > > introduced EtcdKeepAliveService to share one global lease among all > > > > queues > > > > > in a same scheduler.) > > > > > Each ContainerProxy store their information in ETCD with a lease so > > > that > > > > > when a queue tries to create a container, they can easily count the > > > > number > > > > > of existing containers with "Count" API. > > > > > Both data are stored with a lease, if one scheduler or invoker are > > > > failed, > > > > > keep-alive for the given lease is not continued, and finally those > > data > > > > > will be removed. > > > > > > > > > > Follower queues are watching on the leader queue information. If > > there > > > > are > > > > > any changes,(the data will be removed upon scheduler failure) they > > can > > > > > receive the notification and start new leader election. > > > > > When a scheduler is failed, ContainerProxys which were > communicating > > > > with a > > > > > queue in that scheduler, will receive a connection error. > > > > > Then they are designed to access to the ETCD again to figure out > the > > > > > endpoint of the leader queue. > > > > > As one of followers becomes a new leader, ContainerProxys can > connect > > > to > > > > > the new leader. > > > > > > > > > > One thing to note here is, there is only one QueueManager in each > > > > > scheduler. > > > > > One QueueManager holds all queues and delegate requests to the > proper > > > > queue > > > > > in respond to "fetch" requests. > > > > > > > > > > In short, all endpoints data are stored in ETCD and they are > renewed > > > > based > > > > > on keep-alive and lease. > > > > > Each components are designed to access ETCD when the failure > detected > > > and > > > > > connect to a new(failed-over) scheduler. > > > > > > > > > > I hope it is useful to you. > > > > > And I think when I and my colleagues open PRs, we need to add more > > > detail > > > > > design along with them. > > > > > > > > > > If you have any further questions, kindly let me know. > > > > > > > > > > Thanks > > > > > Best regards > > > > > Dominic > > > > > > > > > > > > > > > > > > > > 2019년 4월 6일 (토) 오전 11:28, Mingyu Zhou <[email protected]>님이 작성: > > > > > > > > > > > Dear Dominic, > > > > > > > > > > > > Thanks for your proposal. It is very inspirational and it looks > > > > > promising. > > > > > > > > > > > > I'd like to ask some questions about the fall over/failure > recovery > > > > > > mechanism of the scheduler component. > > > > > > > > > > > > IIUC, a scheduler instance hosts multiple queue managers. If a > > > > scheduler > > > > > is > > > > > > down, we will lose multiple queue managers. Thus, there should be > > > some > > > > > form > > > > > > of failure recovery of queue managers and it raises the following > > > > > > questions: > > > > > > > > > > > > 1. In your proposal, how the failure of a scheduler is detected? > > > I.e., > > > > > > when a scheduler instance is down and some queue manager become > > > > > > unreachable, which component will be aware of this unavailability > > and > > > > > then > > > > > > trigger the recovery procedure? > > > > > > > > > > > > 2. How should the failure be recovered and lost queue managers be > > > > brought > > > > > > back to life? Specifically, in your proposal, you designed a hot > > > > > > standing-by pairing of queue managers (one leader/two followers). > > > Then > > > > > how > > > > > > should the new leader be selected in face of scheduler crash? And > > do > > > we > > > > > > need to allocate a new queue manager to maintain the > > > > > > one-leader-two-follower configuration? > > > > > > > > > > > > 3. How will the other components in the system learn the new > > > > > configuration > > > > > > after a fall over? For example, how will the pool balancer > discover > > > the > > > > > new > > > > > > state of the scheduler it managers and change its policy to > > > distribute > > > > > > queue creation requests? > > > > > > > > > > > > Thanks > > > > > > Mingyu Zhou > > > > > > > > > > > > On Fri, Apr 5, 2019 at 2:56 PM Dominic Kim <[email protected]> > > > > wrote: > > > > > > > > > > > > > Dear David, Matt, and Dascalita. > > > > > > > Thank you for your interest in my proposal. > > > > > > > > > > > > > > Let me answer your questions one by one. > > > > > > > > > > > > > > @David > > > > > > > Yes, I will(and actually already did) implement all components > > > based > > > > on > > > > > > > SPI. > > > > > > > The reason why I said "breaking changes" is, my proposal will > > > affect > > > > > most > > > > > > > of components drastically. > > > > > > > For example, InvokerReactive will become a SPI and current > > > > > > InvokerReactive > > > > > > > will become one of its concrete implementation. > > > > > > > My load balancer and throttler are also based on the current > SPI. > > > > > > > So though my implementation would be included in OpenWhisk, > > > > downstreams > > > > > > > still can take advantage of existing implementation such as > > > > > > > ShardingPoolBalancer. > > > > > > > > > > > > > > Regarding Leader/Follower, a fair point. > > > > > > > The reason why I introduced such a model is to prepare for the > > > future > > > > > > > enhancement. > > > > > > > Actually, I reached a conclusion that memory based activation > > > passing > > > > > > would > > > > > > > be enough for OpenWhisk in terms of message persistence. > > > > > > > But it is just my own opinion and community may want more rigid > > > level > > > > > of > > > > > > > persistence. > > > > > > > I naively thought we can add replication and HA logic in the > > queue > > > > > which > > > > > > > are similar to the one in Kafka. > > > > > > > And Leader/Follower would be a good base building block for > this. > > > > > > > > > > > > > > Currently only a leader fetch activation messages from Kafka. > > > > Followers > > > > > > > will be idle while watching the leadership change. > > > > > > > Once the leadership is changed, one of followers will become a > > new > > > > > leader > > > > > > > and at that time, Kafka consumer for the new leader will be > > > created. > > > > > > > This is to minimize the failure handling time in the aspect of > > > > clients > > > > > as > > > > > > > you mentioned. It is also correct that this flow does not > prevent > > > > > > > activation messages lost on scheduler failure. > > > > > > > But it's not that complex as activation messages are not > > replicated > > > > to > > > > > > > followers and the number of followers are configurable. > > > > > > > If we want, we can configure the number of required queue to 1, > > > there > > > > > > will > > > > > > > be only one leader queue. > > > > > > > (If we ok with the current level of persistence, I think we may > > not > > > > > need > > > > > > > more than 1 queue as you said.) > > > > > > > > > > > > > > Regarding pulling activation messages, each action will have > its > > > own > > > > > > Kafka > > > > > > > topic. > > > > > > > It is same with what I proposed in my previous proposals. > > > > > > > When an action is created, a Kafka topic for the action will be > > > > > created. > > > > > > > So each leader queue(consumer) will fetch activation messages > > from > > > > its > > > > > > own > > > > > > > Kafka topic and there would be no intervention among actions. > > > > > > > > > > > > > > When I and my colleagues open PRs for each component, we will > add > > > > > detail > > > > > > > component design. > > > > > > > It would help you guys understand the proposal more. > > > > > > > > > > > > > > @Matt > > > > > > > Thank you for the suggestion. > > > > > > > If I change the name of it now, it would break the link in this > > > > thread. > > > > > > > I would use the name you suggested when I open a PR. > > > > > > > > > > > > > > > > > > > > > @Dascalita > > > > > > > > > > > > > > Interesting idea. > > > > > > > Any GC patterns do you keep in your mind to apply in OpenWhisk? > > > > > > > > > > > > > > In my proposal, the container GC is similar to what OpenWhisk > > does > > > > > these > > > > > > > days. > > > > > > > Each container will autonomously fetch activations from the > > queue. > > > > > > > Whenever they finish invocation of one activation, they will > > fetch > > > > the > > > > > > next > > > > > > > request and invoke it. > > > > > > > In this sense, we can maximize the container reuse. > > > > > > > > > > > > > > When there is no more activation message, ContainerProxy will > be > > > wait > > > > > for > > > > > > > the given time(configurable) and just stop. > > > > > > > One difference is containers are no more paused, they are just > > > > removed. > > > > > > > Instead of pausing them, containers are waiting for subsequent > > > > requests > > > > > > for > > > > > > > longer time(5~10s) than current implementation. > > > > > > > This is because pausing is also relatively expensive operation > in > > > > > > > comparison to short-running invocation. > > > > > > > > > > > > > > Container lifecycle is managed in this way. > > > > > > > 1. When a container is created, it will add its information in > > > ETCD. > > > > > > > 2. A queue will count the existing number of containers using > > above > > > > > > > information. > > > > > > > 3. Under heavy loads, the queue will request more containers if > > the > > > > > > number > > > > > > > of existing containers is less than its resource limit. > > > > > > > 4. When the container is removed, it will delete its > information > > in > > > > > ETCD. > > > > > > > > > > > > > > > > > > > > > Again, I really appreciate all your feedbacks and questions. > > > > > > > If you have any further questions, kindly let me know. > > > > > > > > > > > > > > Best regards > > > > > > > Dominic > > > > > > > > > > > > > > > > > > > > > > > > > > > > 2019년 4월 5일 (금) 오전 1:24, Dascalita Dragos <[email protected] > >님이 > > > 작성: > > > > > > > > > > > > > > > Hi Dominic, > > > > > > > > Thanks for sharing your ideas. IIUC (and pls keep me honest), > > the > > > > > goal > > > > > > of > > > > > > > > the new design is to improve activation performance. I > > personally > > > > > love > > > > > > > > this; performance is a critical non-functional feature of any > > > FaaS > > > > > > > system. > > > > > > > > > > > > > > > > There’s something I’d like to call out: the management of > > > > containers > > > > > > in a > > > > > > > > FaaS system could be compared to a JVM. A JVM allocates > objects > > > in > > > > > > > memory, > > > > > > > > and GC them. A FaaS system allocates containers to run > actions, > > > and > > > > > it > > > > > > > GCs > > > > > > > > them when they become idle. If we could look at OW's > scheduling > > > > from > > > > > > this > > > > > > > > perspective, we could reuse the proven patterns in the JVM vs > > > > > inventing > > > > > > > > something new. I’d be interested on any GC implications in > the > > > new > > > > > > > design, > > > > > > > > meaning how idle actions get removed, and how is that > > > orchestrated. > > > > > > > > > > > > > > > > Thanks, > > > > > > > > dragos > > > > > > > > > > > > > > > > > > > > > > > > On Thu, Apr 4, 2019 at 8:40 AM Matt Sicker <[email protected] > > > > > > wrote: > > > > > > > > > > > > > > > > > Would it make sense to define an OpenWhisk > > > > Improvement/Enhancement > > > > > > > > > Propoposal or similar that various other Apache projects > do? > > We > > > > > could > > > > > > > > > call them WHIPs or something. :) > > > > > > > > > > > > > > > > > > On Thu, 4 Apr 2019 at 09:09, David P Grove < > > [email protected]> > > > > > > wrote: > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Dominic Kim <[email protected]> wrote on 04/04/2019 > > > 04:37:19 > > > > > AM: > > > > > > > > > > > > > > > > > > > > > > I have proposed a new architecture. > > > > > > > > > > > > > > > > > > > > > > > > > > https://cwiki.apache.org/confluence/display/OPENWHISK/New+architecture > > > > > > > > > > +proposal > > > > > > > > > > > > > > > > > > > > > > It includes many controversial agendas and breaking > > > changes. > > > > > > > > > > > So I would like to form a general consensus on them. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Hi Dominic, > > > > > > > > > > > > > > > > > > > > There's much to like about the proposal. Thank > you > > > for > > > > > > > writing > > > > > > > > > it > > > > > > > > > > up. > > > > > > > > > > > > > > > > > > > > One meta-comment is that the work will have to be > > > done > > > > > in a > > > > > > > way > > > > > > > > > so > > > > > > > > > > there are no actual "breaking changes". It has to be > > > possible > > > > to > > > > > > > > > continue > > > > > > > > > > to configure the system using the existing architectures > > > while > > > > > this > > > > > > > > work > > > > > > > > > > proceeds. I would expect this could be done via a new > > > > > LoadBalancer > > > > > > > and > > > > > > > > > > some deployment options (similar to how Lean OpenWhisk > was > > > > done). > > > > > > If > > > > > > > > > work > > > > > > > > > > needs to be done to generalize the LoadBalancer SPI, that > > > could > > > > > be > > > > > > > done > > > > > > > > > > early in the process. > > > > > > > > > > > > > > > > > > > > On the proposal itself, I wonder if the > complexity > > of > > > > > > > > > Leader/Follower > > > > > > > > > > is actually needed? If a Scheduler crashes, it could be > > > > > restarted > > > > > > > and > > > > > > > > > then > > > > > > > > > > resume handling its assigned load. I think there should > be > > > > > enough > > > > > > > > > > information in etcd for it to recover its current set of > > > > assigned > > > > > > > > > > ContainerProxys and carry on. Activations in its in > > memory > > > > > queues > > > > > > > > would > > > > > > > > > > be lost (bigger blast radius than the current > > architecture), > > > > but > > > > > I > > > > > > > > don't > > > > > > > > > > see that the Leader/Follower changes that (seems way too > > > > > expensive > > > > > > to > > > > > > > > be > > > > > > > > > > replicating every activation in the Follower Queues). > The > > > > > > > > > Leader/Follower > > > > > > > > > > would allow for shorter downtime for those actions > assigned > > > to > > > > > the > > > > > > > > downed > > > > > > > > > > Scheduler, but at the cost of significant complexity. Is > > it > > > > > worth > > > > > > > it? > > > > > > > > > > > > > > > > > > > > Perhaps related to the Leader/Follower, its not > > clear > > > > to > > > > > me > > > > > > > how > > > > > > > > > > activation messages are being pulled from the action > topic > > in > > > > > Kafka > > > > > > > > > during > > > > > > > > > > the Queue creation window. I think they have to go > > somewhere > > > > > > (because > > > > > > > > the > > > > > > > > > > is a mix of actions on a single Kafka topic and we can't > > > stall > > > > > > other > > > > > > > > > > actions while waiting for a Queue to be created for a new > > > > > action), > > > > > > > but > > > > > > > > if > > > > > > > > > > you don't know yet which Scheduler is going to win the > race > > > to > > > > > be a > > > > > > > > > Leader > > > > > > > > > > how do you know where to put them? > > > > > > > > > > > > > > > > > > > > --dave > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > -- > > > > > > > > > Matt Sicker <[email protected]> > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > -- > > > > > > 周明宇 > > > > > > > > > > > > > > > > > > > > >
