I'm using Knative serving with KPA. Autoscaling is available in Knative based on concurrency and RPS. But we need to scale different services based on queue lengths because there are long running async processes. Is there any way we can achieve this in Knative?
I can't use Knative HPA because we need scale to zero feature of Knative.
Thanks in advance!
If you have async (background or scheduled processes), it's likely that Knative is not a good match for your application. There has been some investigation into exposing the HPA v2 custom metrics scanning options (which would might preclude scale to zero, as you note), but even with HPA2 scaling, you'll still run into problems.
The problem with background processes is that Knative and Kubernetes don't have visibility into which Pods are still doing work, so they are equally likely to shut down a Pod doing work as one that is idle.
One workaround would be to move the async work to be synchronous with a request (possibly by using eventing to send a "do work" event), and then processing those events synchronously -- the eventing Broker won't get upset if your requests take a long time to complete. If you're worried about non-uniform processing times, you can even run a second copy of the Knative Service just for handling the long-running requests.
Related
Suppose I want to implement this architecture deployed on Kubernetes cluster:
Gateway
Simple RESTful HTTP microservice accepting scraping tasks (URLs to scrape along with postback urls)
Request Queues - Redis (or other message broker) queues created dynamically per unique domain (when new domain is encountered, gateway should programmatically create new queue. If queue for domain already exists - just place message in it.
Response Queue - Redis (or other message broker) queue used to post Worker results as scraped HTML pages along with postback URLs.
Workers - worker processes which should spin-up at runtime when new queue is created and scale-down to zero when queue is emptied.
Response Workers - worker processes consuming response queue and sending postback results to scraping client. (should be available to scale down to zero).
I would like to deploy the whole solution as dockerized containers on Kubernetes cluster.
So my main concerns/questions would be:
Creating Redis or other message broker queues dynamically at run-time via code. Is it viable? Which broker is best for that purpose? I would prefer Redis if possible since I heard it's the easiest to set up and also it supports massive throughput, ideally my scraping tasks should be short-lived so I think Redis would be okay if possible.
Creating Worker consumers at runtime via code - I need some kind of Kubernetes-compatible technology which would be able to react on newly created queue and spin up Worker consumer container which would listen to that queue and later on would be able to scale up/down based on the load of that queue. Any suggestions for such technology? I've read a bit about KNative, and it's Eventing mechanism, so would it be suited for this use-case? Don't know if I should continue investing my time in reading it's documentation.
Best tools for Redis queue management/Worker management: I would prefer C# and Node.JS tooling. Something like Bull for Node.JS would be sufficient? But ideally I would want to produce queues and messages in Gateway by using C# and consume them in Node.JS (Workers).
If you mean vertical scaling it definitely won't be a viable solution, since it requires pod restarts. Horizontal scaling is somewhat viable when compared to vertical scaling, however you need to consider a fact that even for spinning up your nodes or pods it takes some time and it is always suggested to have proper resources in place for serving your upcoming traffic else this delay will affect some features of your application and there might be a business impact. Just having auto scalers isn’t an option; you should also have proper metrics in place for monitoring your application.
This documentation details how to scale your redis and worker pods respectively using the KEDA mechanism. KEDA stands for Kubernetes Event-driven Autoscaling, KEDA is a plugins which sits on top of existing kubernetes primitives (such as Horizontal pod autoscaler) to scale any number of kubernetes containers based on the number of events which needs to be processed.
Should i engineer my microservice to handle involuntary disruptions like hardware failure?
Are these disruptions frequent enough to be handled in a service running on AWS managed EKS cluster.
Should i consider some design change in the service to handle the unexpected SIGKILL with methods like persisting the data at each step or will that be considered as over-engineering?
What standard way would you suggest for handling these involuntary disruptions if it is
a) a restful service that responds typically in 1s(follows saga pattern).
b) a service that process a big 1GB file in 1 hour.
There are couple of ways to handle those disruptions. As mentioned here here:
Here are some ways to mitigate involuntary disruptions:
Ensure your pod requests the resources it needs.
Replicate your application if you need higher availability. (Learn about running replicated stateless and stateful applications.)
For even higher availability when running replicated applications, spread applications across racks (using anti-affinity) or across zones
(if using a multi-zone cluster.)
The frequency of voluntary disruptions varies.
So:
if your budget allows it, spread your app accross zones or racks, you can use Node affinity to schedule Pods on cetrain nodes,
make sure to configure Replicas, it will ensure that when one Pod receives SIGKILL the load is automatically directed to another Pod. You can read more about this here.
consider using DaemonSets, which ensure each Node runs a copy of a Pod.
use Deployments for stateless apps and StatefulSets for stateful.
last thing you can do is to write your app to be distruption tolerant.
I hope I cleared the water a little bit for you, feel free to ask more questions.
I have an (internal) K8s deployment (Python, TensorFlow, Guinicorn) with around 60 replicas and an attached K8s service to distribute incoming HTTP requests. Each of these 60 pods can only actually process one HTTP request at a time (because of TensorFlow reasons). Processing a request takes between 1 and 4 seconds. If a second request is sent to a pod while it's still processing one, the second request is just queued in the Gunicorn backlog.
Now I'd like to reduce the probability of that queuing happening as much as possible, i.e., have new requests routed to one of the non-occupied pods as long as such a non-occupied one exists.
Round-robin would not do the trick, because not every request takes the same amount of time to answer (see above).
The Python application itself could make the endpoint used for the ReadinessProbe fail while it's processing a normal request, but as far as I understand, readiness probes are not meant for something that dynamic (K8s would need to poll them multiple times per second).
So how could I achieve the goal?
Can't you implement the pub/sub or message broker in between?
saver data into a queue based on the ability you worker will fetch the message or data from queue and request will get processed.
You can use Redis for creating queues and in queue, you can use pub/sub also possible using the library. i used one before in Node JS however could be possible to implement the same using python also.
in 60 replicas ideally worker or we can say scriber will be running.
As soon as you get a request one application will publish it and scribers will be continuously working for processing those messages.
We also implemented one step further, scaling the worker count automatically depends on the message count in the queue.
This library i am using with the Node js : https://github.com/OptimalBits/bull
...kubectl get service shows "TYPE ClusterIP" and "EXTERNAL-IP <none>
Your k8s service will be routing requests at random in this case... obviously not good to your app. If you would like to stick with kube-proxy, you can switch to ipvs mode with sed. Here's a good article about it. Otherwise, you can consider using some sort of ingress controller like the one mentioned earlier on; ingress-nginx with "ewma" mode.
This is a purely theoretical question. A standard Kubernetes clusted is given with autoscaling in place. If memory goes above a certain targetMemUtilizationPercentage than a new pod is started and it takes on the flow of requests that is coming to the contained service. The number of minReplicas is set to 1 and the number of maxReplicas is set to 5.
What happens when the number of pods that are online reaches maximum (5 in our case) and requests from clients are still coming towards the node? Are these requests buffered somewhere of they are discarded? Can I take any actions to avoid request loss?
Natively Kubernetes does not support messaging queue buffering. Depends on the scenario and setup you use your requests will most likely 'timeout'. To efficiently manage those you`ll need custom resource running inside Kubernetes cluster.
In that situations it very common to use a message broker which ensures communication between microservices is reliable and stable, that the messages are managed and monitored within the system and that messages don’t get lost.
RabbitMQ, Kafka and Redis appears to be most popular but choosing the right one will heaving depend on your requirement and features needed.
Worth to note since Kubernetes essentially runs on linux is that linux itself also manages/limits the requests coming in socket. You may want to read more about it here.
Another thing is that if you have pods limits set or lack of resource it is most likely that pods might be restarted or cluster will become unstable. Usually you can prevent it by configuring some kind of "circuit breaker" to limit amount of requests that could go to backed without overloading it. If the amount of requests goes beyond the circuit breaker threshold, excessive requests will be dropped.
It is better to drop some request than having cascading failure.
I managed to test this scenario and I get 503 Service Unavailable and 403 Forbidden on my requests that do not get processed.
Knative Serving actually does exactly this. https://github.com/knative/serving/
It buffers requests and informs autoscaling decisions based on in-flight request counts. It also can enforce per-Pod max in-flight requests and hold onto request until newly scaled-up Pods come up and then Knative proxies the request to them as it has this container named queue-proxy as a sidecar to its workload type called "Service".
I'm running a fairly resource-intensive service on a Kubernetes cluster to support CI activities. Only a single replica is needed, but it uses a lot of resources (16 cpu), and it's only needed during work hours generally (weekdays, 8am-6pm roughly). My cluster runs in a cloud and is setup with instance autoscaling, so if this service is scaled to zero, that instance can be terminated.
The service is third-party code that cannot be modified (well, not easily). It's a fairly typical HTTP service other than that its work is fairly CPU intensive.
What options exist to automatically scale this Deployment down to zero when idle?
I'd rather not setup a schedule to scale it up/down during working hours because occasionally CI activities are performed outside of the normal hours. I'd like the scaling to be dynamic (for example, scale to zero when idle for >30 minutes, or scale to one when an incoming connection arrives).
Actually Kubernetes supports the scaling to zero only by means of an API call, since the Horizontal Pod Autoscaler does support scaling down to 1 replica only.
Anyway there are a few Operator which allow you to overtake that limitation by intercepting the requests coming to your pods or by inspecting some metrics.
You can take a look at Knative or Keda.
They enable your application to be serverless and they do so in different ways.
Knative, by means of Istio intercept the requests and if there's an active pod serving them, it redirects the incoming request to that one, otherwise it trigger a scaling.
By contrast, Keda best fits event-driven architecture, because it is able to inspect predefined metrics, such as lag, queue lenght or custom metrics (collected from Prometheus, for example) and trigger the scaling.
Both support scale to zero in case predefined conditions are met in a equally predefined window.
Hope it helped.
I ended up implementing a custom solution: https://github.com/greenkeytech/zero-pod-autoscaler
Compared to Knative, it's more of a "toy" project, fairly small, and has no dependency on istio. It's been working well for my use case, though I do not recommend others use it without being willing to adopt the code as your own.
There are a few ways this can be achieved, possibly the most "native" way is using Knative with Istio. Kubernetes by default allows you to scale to zero, however you need something that can broker the scale-up events based on an "input event", essentially something that supports an event driven architecture.
You can take a look at the offcial documents here: https://knative.dev/docs/serving/configuring-autoscaling/
The horizontal pod autoscaler currently doesn’t allow setting the minReplicas field to 0, so the autoscaler will never scale down to zero, even if the pods aren’t doing anything. Allowing the number of pods to be scaled down to zero can dramatically increase the utilization of your hardware.
When you run services that get requests only once every few hours or even days, it doesn’t make sense to have them running all the time, eating up resources that could be used by other pods.
But you still want to have those services available immediately when a client request comes in.
This is known as idling and un-idling. It allows pods that provide a certain service to be scaled down to zero. When a new request comes in, the request is blocked until the pod is brought up and then the request is finally forwarded to the pod.
Kubernetes currently doesn’t provide this feature yet, but it will eventually.
based on documentation it does not support minReplicas=0 so far. read this thread :-https://github.com/kubernetes/kubernetes/issues/69687. and to setup HPA properly you can use this formula to setup required pod :-
desiredReplicas = ceil[currentReplicas * ( currentMetricValue / desiredMetricValue )]
you can also setup HPA based on prometheus metrics follow this link:-
https://itnext.io/horizontal-pod-autoscale-with-custom-metrics-8cb13e9d475