RFC: single binary cortex / make cortex easier to get started with

[tomwilkie]

• I'd like to put Cortex back to a single binary; see Split up cortex into different binaries / images #151 for the split into multiple binaries.
• You'd still be able to run as a ingester, distributor etc; that would be a flag.
• I want to introduce reading flags from yaml (see Underpinnings of a better single-process, no-dependancies Cortex. #1127); this would allow us to leave the flag defaults as is for backwards compat, whilst shipping a couple of yamls in the repo for getting started.
• I want to unify the dependancies of the various "components" into a single place, and programatically initialise them in the right order. This should reduce the duplication in the various mains.
• it will ensure the lite version never breaks due to flag (although we have a test now, so maybe not an issue)

I've got a WIP prototype of this based on the code in #1127, which I'm trying to publish in time for kubecon. WDYT?

[bboreham]

Would this apply to the query front-end too? It seems a different style to the ingester, querier, etc.

I see some other pros:

• faster build time

and some cons:

• all metrics will be registered in all components (hence appear as all-zero timeseries)
• all the flags will come out when it prints usage

[tomwilkie]

Would this apply to the query front-end too? It seems a different style to the ingester, querier, etc.

I'd like it to. Its more like the ruler in that its a singleton; in this mode we'd need some was of doing master-election to pick which lite instance would be a query-frontend. Not insurmountable though.

all the flags will come out when it prints usage

True; I suspect when we get closer to v1 and we're more comfortable with the config files, we could consider not exposing some of them as flags?