Merge pull request #2636 from wojtek-t/network_latency

Network latency SLI

sig-scalability/slos/network_latency.md

@@ -0,0 +1,54 @@
+## In-cluster network latency SLIs/SLOs details
+
+### Definition
+
+| Status | SLI | SLO |
+| --- | --- | --- |
+| __WIP__ | In-cluster network latency from a single prober pod, measured as latency of per second ping from that pod to "null service", measured as 99th percentile over last 5 minutes. | In default Kubernetes installataion with RTT between nodes <= Y, 99th percentile of (99th percentile over all prober pods) per cluster-day <= X |
+
+### User stories
+- As a user of vanilla Kubernetes, I want some guarantee how fast my http
+request to some Kubernetes service reaches its endpoint
+
+### Other notes
+- We obviously can't give any guarantee in a general case, because cluster
+administrators may configure cluster as they want.
+- As a result, we define the SLI to be very generic (no matter how your cluster
+is set up), but we provide SLO only for default installations with an additional
+requirement that low-level RTT between nodes is lower than Y.
+- Network latency is one of the most crucial aspects from the point of view
+of application performance, especially in microservices world. As a result, to
+meet user expectations, we need to provide some guarantees arount that.
+- We decided for the SLI definition as formulated above, because:
+  - it represents a user oriented end-to-end flow - it involves among others
+    latency of in-cluster network programming mechanism (e.g. iptables). <br/>
+    __TODO:__ We considered making DNS resolution part of it, but decided not
+    to mix them. However, longer term we should consider joining them.
+  - it is easily measurable in all running clusters in which we can run probers
+    (e.g. measuring request latencies coming from all pods on a given
+    node would require some additional instrumentation, such as a side car for
+    each of them, and that overhead may be not acceptable in many cases)
+  - it is not application-specific
+
+### Caveats
+- The SLI is formulated for a prober pods, even though users are mostly
+interested in the aggregation across all pods (that is done only at the SLO
+level). However, that provides very similar guarantees and makes it fairly
+easy to measure.
+- The RTT between nodes may significantly differ, if nodes are in different
+topologies (e.g. GCP zones). However, given that topology-aware service routing
+is not natively supported in Kubernetes yet, we explicitly acknowledge that
+depending on the pinged endpoint, results may signiifcantly differ if nodes
+are spanning multiple topologies.
+- The prober reporting that is fairly trivial and itself needs only negligible
+amount of resources. Unfortunately there isn't any component to which we can
+attach that functionality (e.g. KubeProxy is running in host network), so
+**we will create a dedicated set of prober pods**. We will run a set of prober
+pods (number proportional to cluster size).
+- We don't have any "null service" running in cluster, so an administrator has
+to set up one to make the SLI measurable in real cluster. In tests, we will
+create a service on top of prober pods.
+
+### Test scenario
+
+__TODO: Describe test scenario.__

sig-scalability/slos/slos.md

@@ -108,6 +108,7 @@ Prerequisite: Kubernetes cluster is available and serving.
 | __Official__ | Startup latency of stateless and schedulable pods, excluding time to pull images and run init containers, measured from pod creation timestamp to when all its containers are reported as started and observed via watch, measured as 99th percentile over last 5 minutes | In default Kubernetes installation, 99th percentile per cluster-day<sup>[1](#footnote1)</sup> <= 5s | [Details](./pod_startup_latency.md) |
 | __WIP__ | Latency of programming a single (e.g. iptables on a given node) in-cluster load balancing mechanism, measured from when service spec or list of its `Ready` pods change to when it is reflected in load balancing mechanism, measured as 99th percentile over last 5 minutes | In default Kubernetes installation, 99th percentile of (99th percentiles across all programmers (e.g. iptables)) per cluster-day<sup>[1](#footnote1)</sup> <= X | [Details](./network_programming_latency.md) |
 | __WIP__ | Latency of programming a single in-cluster dns instance, measured from when service spec or list of its `Ready` pods change to when it is reflected in that dns instance, measured as 99th percentile over last 5 minutes | In default Kubernetes installation, 99th percentile of (99th percentiles across all dns instances) per cluster-day <= X | [Details](./dns_programming_latency.md) |
+| __WIP__ | In-cluster network latency from a single prober pod, measured as latency of per second ping from that pod to "null service", measured as 99th percentile over last 5 minutes. | In default Kubernetes installataion with RTT between nodes <= Y, 99th percentile of (99th percentile over all prober pods) per cluster-day <= X | [Details](./network_latency.md) |
 
 <a name="footnote1">\[1\]</a> For the purpose of visualization it will be a
 sliding window. However, for the purpose of reporting the SLO, it means one