Add Metrics for Cache

[alpe]

It would be great if we could use prometheus to collect metrics for the cache.
Useful metrics would be

• hits/misses
• memory used/ available

While there is a Rust lib for prometheus the integration into the metrics server (Go side) can be a bit complicated (wrapped in cosmos-sdk/telemetry, which uses github.com/armon/go-metrics/prometheus.

A "simpler" solution may be adding a return value for the Go calls that contains cache hit/memory-used data that is then fed into prometheus on the Go side.

[webmaster128]

There is a stats call allready, returning

#[derive(Debug, Default, Clone, Copy)]
pub struct Stats {
    pub hits_memory_cache: u32,
    pub hits_fs_cache: u32,
    pub misses: u32,
}

This seems like a good start to me. It can easily be extended. This is not yet connected into Go in wasmvm, but I'd rather do that than adding prometheus here.

[webmaster128]

Hey @alpe, in #877 I started exposing a Metrics object that contains (when flattened):

struct Metrics {
    pub hits_pinned_memory_cache: u32,
    pub hits_memory_cache: u32,
    pub hits_fs_cache: u32,
    pub misses: u32,
    pub elements_pinned_memory_cache: usize,
    pub elements_memory_cache: usize,
    pub size_pinned_memory_cache: usize,
    pub size_memory_cache: usize,
}

Is this a useful set of data to to start with? I especially wonder about hits/misses, which is a a sum that accumulates over time. Can this be processed nicely? Or should we calculate always current valus (like moving average of hit ration)?

[alpe]

Is this a useful set of data to to start with? I especially wonder about hits/misses, which is a a sum that accumulates over time. Can this be processed nicely? Or should we calculate always current valus (like moving average of hit ration)?

This looks good. Counter or Gauge (incr/decr) types are fine. We can always do calculation, windows, ... with them in prometheus. Better keep them simple here.

Does it make sense to return the available memory cache? If it always fills up completely then it is not relevant, I guess.

[webmaster128]

Does it make sense to return the available memory cache? If it always fills up completely then it is not relevant, I guess.

The value you are looking for is available indirectly as cache size (configurable per node) - metrics.size_memory_cache. I.e. we report the usage, not the available memory. Let's start with that.