Dynamic fingerprints made the code much slower, pprof shows they are responsible for ~70% of all cpu usage for any API call. To make it worse they are applied to all alerts, since dedup layer doesn't know which alerts will be filtered later, it operates on all of them. This PR will:
1. add benchmarks to so it's easier to track performance
2. Keep current methods for accessing fingerprints, but use precomputed static fields in those
3. Refactor Alert methods to use pointers, so we're not working on a copy
Benchmark timing went down from ~4000ns to 0.4ns for fingerprint calls and API response times from 1.3s (for my test sample) to 0.2s, which puts it back to the same level as before moving fingerprints to be dynamic. I still don't like this code much, it's all over the place, but I don't have a good idea how to better structure this, let's hope I'll be wiser in the future.
Fixes#104
@status filter was added to the master branch to support new status key from Alertmanager >=0.6.1
status ended up being nested in Alertmanager (it was added to solve AM issue 609 and that was a long PR with lots of changes), current unsee implementation ended being slightly off with how Alertmanager is naming this, it should actually be @state rather than @status.
Expose counts of all alert statuses in each group (active, suppressed, unprocessed) instead of only active, and if a group contains only unprocessed alerts then color it as gray
Now that Alertmanager provides status=(unprocessed|suppressed|active) in the API it doesn't make sense to keep having separate filters for each value, merge @silenced & @inhibited into a single filter, update docs and UI
This will make the UI present inhibited alerts same way as silenced. It also drops unused fields in the model, replacing SilencedCount & UnsilencedCount with only ActiveCount
