Marco Pracucci

Thanks @ortuman for working on this! Working on your personal repo is fine for experimenting, but if we'll end up vendoring this in Mimir then please fork `gomemcached` under the...

I would like to better understand why a very quick propagation is required for the HA tracker. Given the HA tracker triggers a failover once we're not receiving any sample...

> For example if one distributor doesn't see up-to-date received_at, and decides to make a switch to different replica With the default config, the `received_at` is updated every 15s while...

> I think we may be able to get reliable state replication with memberlist (

> Can we use ingesters somehow to store this data? Dimitar mentioned it too in a private conversation. I think we're just giving for granted memberlist wouldn't work, but before...

> what happens when we increase the size of messages? so far I was sending a unix millisecond timestamp as a string; that's 13 bytes and is unrealistic For the...

> > Does CPU and memory consumption scale linearly with cluster size? > I would expect it to do so, right? Every extra client can potentially send you a gossip...

> alertmanager FYI, Mimir jsonnet doesn't support alertmanager deployment in multi-zone. If the primary purposes of multi-zone is to speed up rollouts, then it's non issue for alertmanagers (you typically...

> Documenting the migration strategy for existing Helm users: _I'm leaving few minor comments below, just to be on the same page._ > Step 1 > Deploy 3 new StatefulSets...

> > The problem is that when you'll revert the change back (which requires an ingesters rollout), it will unregister ingesters and flush blocks for multi-zone ingesters too. > >...