package balancergroup

import "google.golang.org/grpc/xds/internal/balancer/balancergroup"

Package balancergroup implements a utility struct to bind multiple balancers into one balancer.

Index ¶

• Variables
• type BalancerGroup
• func New(cc balancer.ClientConn, bOpts balancer.BuildOptions, stateAggregator BalancerStateAggregator, loadStore load.PerClusterReporter, logger *grpclog.PrefixLogger) *BalancerGroup
• func (bg *BalancerGroup) Add(id string, builder balancer.Builder)
• func (bg *BalancerGroup) Close()
• func (bg *BalancerGroup) ExitIdle()
• func (bg *BalancerGroup) Remove(id string)
• func (bg *BalancerGroup) ResolverError(err error)
• func (bg *BalancerGroup) Start()
• func (bg *BalancerGroup) UpdateClientConnState(id string, s balancer.ClientConnState) error
• func (bg *BalancerGroup) UpdateSubConnState(sc balancer.SubConn, state balancer.SubConnState)
• type BalancerStateAggregator

Variables ¶

var DefaultSubBalancerCloseTimeout = 15 * time.Minute

DefaultSubBalancerCloseTimeout is defined as a variable instead of const for testing.

TODO: make it a parameter for New().

Types ¶

type BalancerGroup ¶

type BalancerGroup struct {
	// contains filtered or unexported fields
}

BalancerGroup takes a list of balancers, and make them into one balancer.

Note that this struct doesn't implement balancer.Balancer, because it's not intended to be used directly as a balancer. It's expected to be used as a sub-balancer manager by a high level balancer.

Updates from ClientConn are forwarded to sub-balancers

• service config update
• Not implemented
• address update
• subConn state change
• find the corresponding balancer and forward

Actions from sub-balances are forwarded to parent ClientConn

• new/remove SubConn
• picker update and health states change
• sub-pickers are sent to an aggregator provided by the parent, which will group them into a group-picker. The aggregated connectivity state is also handled by the aggregator.
• resolveNow

Sub-balancers are only built when the balancer group is started. If the balancer group is closed, the sub-balancers are also closed. And it's guaranteed that no updates will be sent to parent ClientConn from a closed balancer group.

func New ¶

func New(cc balancer.ClientConn, bOpts balancer.BuildOptions, stateAggregator BalancerStateAggregator, loadStore load.PerClusterReporter, logger *grpclog.PrefixLogger) *BalancerGroup

New creates a new BalancerGroup. Note that the BalancerGroup needs to be started to work.

TODO(easwars): Pass an options struct instead of N args.

func (*BalancerGroup) Add ¶

func (bg *BalancerGroup) Add(id string, builder balancer.Builder)

Add adds a balancer built by builder to the group, with given id.

func (*BalancerGroup) Close ¶

func (bg *BalancerGroup) Close()

Close closes the balancer. It stops sub-balancers, and removes the subconns. The BalancerGroup can be restarted later.

func (*BalancerGroup) ExitIdle ¶

func (bg *BalancerGroup) ExitIdle()

ExitIdle should be invoked when the parent LB policy's ExitIdle is invoked. It will trigger this on all sub-balancers, or reconnect their subconns if not supported.

func (*BalancerGroup) Remove ¶

func (bg *BalancerGroup) Remove(id string)

Remove removes the balancer with id from the group.

But doesn't close the balancer. The balancer is kept in a cache, and will be closed after timeout. Cleanup work (closing sub-balancer and removing subconns) will be done after timeout.

func (*BalancerGroup) ResolverError ¶

func (bg *BalancerGroup) ResolverError(err error)

ResolverError forwards resolver errors to all sub-balancers.

func (*BalancerGroup) Start ¶

func (bg *BalancerGroup) Start()

Start starts the balancer group, including building all the sub-balancers, and send the existing addresses to them.

A BalancerGroup can be closed and started later. When a BalancerGroup is closed, it can still receive address updates, which will be applied when restarted.

func (*BalancerGroup) UpdateClientConnState ¶

func (bg *BalancerGroup) UpdateClientConnState(id string, s balancer.ClientConnState) error

UpdateClientConnState handles ClientState (including balancer config and addresses) from resolver. It finds the balancer and forwards the update.

func (*BalancerGroup) UpdateSubConnState ¶

func (bg *BalancerGroup) UpdateSubConnState(sc balancer.SubConn, state balancer.SubConnState)

UpdateSubConnState handles the state for the subconn. It finds the corresponding balancer and forwards the update.

type BalancerStateAggregator ¶

type BalancerStateAggregator interface {
	// UpdateState updates the state of the id.
	//
	// It's up to the implementation whether this will trigger an update to the
	// parent ClientConn.
	UpdateState(id string, state balancer.State)
}

BalancerStateAggregator aggregates sub-picker and connectivity states into a state.

It takes care of merging sub-picker into one picker. The picking config is passed directly from the the parent to the aggregator implementation (instead via balancer group).

Source Files ¶

balancergroup.go balancerstateaggregator.go