healthz?verbose can help us here, if we can get a curl off before a crashloop. in this example, we can see etcd is down. And because it's down, apiserver fails its own healthcheck and crashloops.

the default storage limit for etcd is 2gb - so etcd_object_count can be a clue to an overwhelmed etcd

etcd objects aren't all the same size. a watch mechanism means etcd has to store historical versions of an object. So a very frequently updated object can take up a disproportionate amount of space on etcd!

etcd compacts every 5 minutes, so you can easily go over storage limit sizes. you can use kubectl for this if apiserver is up, but if it's not this is where auger shines!

probably don't store images in etcd, btw. it's not meant for that

last example: slow API server! let's take a look at request latency. it turned out these metrics were useless, because the bucket sizes were too coarse.

bad metrics make it difficult for cluster operators. but updating is hard! Changing metrics could cause false positives or negatives in people's monitoring

bad metrics can't be disabled, so you need a full upgrade. sig-instrumentation did an overhaul of broken metrics in 1.14.
- labels didn't match
- wrong data types
- units weren't standardized, and many were even outright incorrect!

how do we deprecate bad metrics? sig-instrumentation is going to treat metrics as a proper api. we'll have stability levels for all metrics going forward, so we can mark metrics as deprecated

what metrics are stable? we need criteria and promotion still. and we need runtime flags for disabling individual metrics

distributed tracing and structured logs are (hopefully) coming soon too!

(that's all for today. Thanks @ehashdn and @LogicalHan!) #KubeCon