Skip to content

Latest commit

 

History

History
165 lines (123 loc) · 6.55 KB

201807_store_instance_high_availability.md

File metadata and controls

165 lines (123 loc) · 6.55 KB
title type menu status owner
High-availability for store instances
proposal
proposals
rejected
mattbostock

Summary

This proposal makes total sense and solves our goals when using gossip. However there exists a very easy solution to this problem in form of using just static entry with any loadbalancer like Kubernetes Service to load balance through different Store Gateways. Those are technically stateless, so request can fetch the data independently.

Motivation

Thanos store instances currently have no explicit support for high-availability; query instances treat all store instances equally. If multiple store instances are used as gateways to a single bucket in an object store, Thanos query instances will wait for all instances to respond (subject to timeouts) before returning a response.

Goals

  • Explicitly support and document high availability for store instances.

  • Reduce the query latency incurred by failing store instances when other store instances could return the same response faster.

Proposal

Thanos supports deduplication of metrics retrieved from multiple Prometheus servers to avoid gaps in query responses where a single Prometheus server failed but similar data was recorded by another Prometheus server in the same failure domain. To support deduplication, Thanos must wait for all Thanos sidecar servers to return their data (subject to timeouts) before returning a response to a client.

When retrieving data from Thanos bucket store instances, however, the desired behaviour is different; we want Thanos use the first successful response it receives, on the assumption that all bucket store instances that communicate with the same bucket have access to the same data.

To support the desired behaviour for bucket store instances while still allowing for deduplication, we propose to expand the InfoResponse Protobuf used by the Store API by adding two fields:

For example;

--- before	2018-07-02 15:49:09.000000000 +0100
+++ after	2018-07-02 15:49:13.000000000 +0100
@@ -1,5 +1,6 @@
 message InfoResponse {
   repeated Label labels = 1 [(gogoproto.nullable) = false];
   int64 min_time        = 2;
   int64 max_time        = 3;
+  string store_group_id = 4;
+  enum PeerType {
+    STORE  = 0;
+    SOURCE = 1;
+    QUERY  = 2;
+  }
+  PeerType peer_type    = 5;
 }

For the purpose of querying data from store instances, stores instance will be grouped by:

  • labels, as returned as part of InfoResponse
  • the new store_group_id string identifier

Therefore, stores having identical sets of labels and identical values for store_group_id will belong in the same group for the purpose of querying data. Stores having an empty store_group_id field and matching labels will be considered to be part of the same group. Stores having an empty store_group_id field and empty label sets will also be considered part of the same group.

If a service implementing the store API (a 'store instance') has a STORE or QUERY peer type, query instances will treat each store instance in the same group as having access to the same data. Query instances will randomly pick any two store instances1 from the same group and use the first response returned.

Otherwise, for the SOURCE peer type, query instances will wait for all instances within the same group to respond (subject to existing timeouts) before returning a response, consistent with the current behaviour. This is necessary to collect all data available for the purposes of deduplication and to fill gaps in data where an individual Prometheus server failed to ingest data for a period of time.

Each service implementing the store API must determine what value the store_group_id should return. For bucket stores, store_group_id should contain the concatenation of the object store URL and bucket name. For all other existing services implementing the store API, we will use an empty string for store_group_id until a reason exists to use it.

Multiple buckets or object stores will be supported by setting the store_group_id.

Existing instances running older versions of Thanos will be assumed to have an empty string for store_group_id and a SOURCE peer type, which will retain existing behaviour when awaiting responses.

Scope

Horizontal scaling should be handled separately and is out of scope for this proposal.

User experience

From a user's point of view, query responses should be faster and more reliable:

  • Running multiple bucket store instances will allow the query to be served even if a single store instance fails.

  • Query latency should be lower since the response will be served from the first bucket store instance to reply.

The user experience for query responses involving only Thanos sidecars will be unaffected.

Alternatives considered

Implicitly relying on store labels

Rather than expanding the InfoResponse Protobuf, we had originally considered relying on an empty set of store labels to determine that a store instance was acting as a gateway.

We decided against this approach as it would make debugging harder due to its implicit nature, and is likely to cause bugs in future.

Using boolean fields to determine query behaviour

We rejected the idea of adding a gateway or deduplicated boolean field to InfoResponse in the store RPC API. The value of these fields would have had the same effect on query behaviour as returning the peer type field as proposed above and would be more explicit, but were specific to this use case.

The peer type field in InfoResponse proposed above could be used for other use cases aside from determining query behaviour.

Related future work

Sharing data between store instances

Thanos bucket stores download index and metadata from the object store on start-up. If multiple instances of a bucket store are used to provide high availability, each instance will download the same files for its own use. These file sizes can be in the order of gigabytes.

Ideally, the overhead of each store instance downloading its own data would be avoided. We decided that it would be more appropriate to tackle sharing data as part of future work to support the horizontal scaling of store instances.