Skip to content

Latest commit

 

History

History
62 lines (42 loc) · 4.77 KB

201809-gossip-removal.md

File metadata and controls

62 lines (42 loc) · 4.77 KB
type title status owner menu
proposal
Deprecated gossip clustering in favor of File SD
completed
bwplotka
proposals-done

Summary

It is becoming clear that we need to remove gossip protocol as our main way of communication between Thanos Querier and other components. Static configuration seems to be well enough for our simple use cases. To give users more flexibility (similar to gossip auto-join logic), we already wanted to introduce a File SD that allows changing StoreAPIs on-the-fly.

Motivation

Gossip protocol (with the membership implementation) was built into Thanos from the very beginning. The main advantages over other solution to connect components were:

  • Auto-join and auto-drop of components based on health checks.
  • Propagation of tiny metadata.

After a couple of month of maintaining Thanos project and various discussions with different users, we realized that those advantages are not outstanding anymore and are not worth keeping, compared to the issues gossip causes. There are numerous reasons why we should deprecate gossip:

  • Gossip has been proven to be extremely confusing for the new users. Peer logic and really confusing cluster.advertise-address that was sometimes able to magically deduce private IP address (and sometimes not!) were leading to lots of questions and issues. Something that was made for quick ramp-up into Thanos ("just click the button and it auto-joins everything") created lots of confusion and made it harder to experiment.
  • Auto-join logic has its price. All peers connected were aware of each other. All were included in metadata propagation. Membership project community made an awesome job to optimize N^N communication (all-to-all), but nevertheless, it is something that is totally unnecessary to Thanos. And it led to wrong assumptions that e.g sidecar needs to be aware of another sidecar etc. Thanos Querier is the only component that requires to be aware of other StoreAPIs. It is clear that all-to-all communication is neither necessary nor educative.
  • Unless specifically configured (which requires advanced knowledge) Gossip uses mix of TCP and UPD underneath its custom app level protocol. This is a no-go if you use L7 loadbalancers and proxies.
  • Global gossip is really difficult to achieve. BTW, If you know how to setup this, please write a blog about it! (:
  • With the addition of the simplest possible solution to give Thanos Querier knowledge where are StoreAPIs (static --store flag), we needed to implement health check and metadata propagation anyway. In fact, StoreAPI.Info was already there all the time.
  • Gossip operates per peer level and there is no way you can abstract multiple peers behind loadbalancer. This hides easy solutions from our eyes, e.g how to make Store Gateway HA. Without gossip, you can just use Kubernetes HA Service or any other loadbalancer. To support Store Gateway HA for gossip we would end up implementing LB logic in Thanos Querier (like proposed here)
  • At some point we want to be flexible and allow other discovery mechanisms. Gossip does not work for everyone and static flags are too.. static. (: We need File SD for flexibility anyway.
  • One thing less to maintain.

Goals

  • Remove gossip support from code (Decision after initial feedback)
    • We are still RC, so technically there are no API guarantees yet.
  • Leave --store flags
  • Make sure File SD is in place and documented before removal.

Proposal: Steps

  • Add File Service Discovery (SD): thanos-io#492
  • Remove gossip from the documentation, be clear what talks with what (!)
  • Deprecate gossip in code.
  • Remove gossip code and flags AFTER File SD is done and stable.

Backwards compatibility

We will break backward compatibility, as gossip was the main way to connect components. As it is not very nice, it will ensure stability and better user experience and flexibility in the future. As we are in 0.1.0 RC process, we will not break any guarantee.

As help for migration we might want to add sidecar that uses FileSD but support gossip clustering.

Alternatives considered

  1. Just improve documentation of gossip
  • It will not help in efficiency/unnecessary traffic problem
  • We will have 2 ways of communication as it is now - hard to maintain and think about.

Related Future Work

  1. Add example sidecars that will provide more sophisticated discovery and integrate with our File SD (e.g k8s SD)