Refactoring ClusterManagerImpl to support dynamic config for composite filter when its in upstream filter chain

yanjunxiang-google commented 7 months ago

Title: Refactoring ClusterManagerImpl c'tor and to support dynamic config for composite filter when its in upstream filter chain

This issue is related to the PR: https://github.com/envoyproxy/envoy/pull/33013

Description: Currently, the cluster manager is initialized with the below sequence:

1) https://github.com/envoyproxy/envoy/blob/8ff28fe73fe0c704b416c790fe7b22622cd5d5fc/source/server/configuration_impl.cc#L131 2)

https://github.com/envoyproxy/envoy/blob/8ff28fe73fe0c704b416c790fe7b22622cd5d5fc/source/common/upstream/cluster_manager_impl.cc#L2227

i.e, it calls the cluster_manager_impl->init(bootstrap) before return unique_ptr, and assign to the clustermanager.

This is causing problems when functions called inside cluster_manager_impl->init(bootstrap) needs to access clustermanger, which is the case for composite filter in the upstream filter chain.

So, for composite filter case, if it is put in the upstream filter chain, when the dynamic_config is configured: https://github.com/envoyproxy/envoy/blob/8ff28fe73fe0c704b416c790fe7b22622cd5d5fc/api/envoy/extensions/filters/http/composite/v3/composite.proto#L59

the above cluster_manager_impl->init(bootstrap) will eventually call this function: https://github.com/envoyproxy/envoy/blob/8ff28fe73fe0c704b416c790fe7b22622cd5d5fc/source/extensions/filters/http/composite/action.cc#L43 which access above clustermanger through server_factory_context.clusterManager(). As clustermanager is nullptr, this crashed.

So, if we want to support composite filter with dynamic_config in the upstream filter chain, one of the solution is moving the cluster_manager_impl->init(bootstrap) out from ProdClusterManagerFactory::clusterManagerFromProto

And call it after, like this:

clustermanager = clusterManagerFromProto(); clustermanager->init(bootstrap);

This way, when the above init() function access clusterManager(), clustermanager is already setup and no issues. We need to add good comments to make sure no other code wrote in between these two lines though.

[optional Relevant Links:]