-
Notifications
You must be signed in to change notification settings - Fork 107
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Merge pull request #5528 from oasisprotocol/kostko/feature/rt-load-ba…
…lancing go/runtime: Add load-balancer runtime provisioner
- Loading branch information
Showing
15 changed files
with
412 additions
and
110 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1 @@ | ||
| go/runtime: Add load-balancer runtime provisioner |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,300 @@ | ||
| // Package loadbalance implements a runtime provisioner that internally load-balances requests among | ||
| // multiple runtime instances. This is especially useful on client nodes handling queries. | ||
| package loadbalance | ||
|
|
||
| import ( | ||
| "context" | ||
| "errors" | ||
| "fmt" | ||
| "sync" | ||
|
|
||
| "github.com/prometheus/client_golang/prometheus" | ||
|
|
||
| "github.com/oasisprotocol/oasis-core/go/common" | ||
| "github.com/oasisprotocol/oasis-core/go/common/logging" | ||
| "github.com/oasisprotocol/oasis-core/go/common/node" | ||
| "github.com/oasisprotocol/oasis-core/go/common/pubsub" | ||
| "github.com/oasisprotocol/oasis-core/go/runtime/host" | ||
| "github.com/oasisprotocol/oasis-core/go/runtime/host/protocol" | ||
| ) | ||
|
|
||
| // Config is the configuration for the runtime load balancer. | ||
| type Config struct { | ||
| // NumInstances is the number of runtime instances to provision. | ||
| NumInstances int | ||
| } | ||
|
|
||
| type lbRuntime struct { | ||
| id common.Namespace | ||
| instances []host.Runtime | ||
|
|
||
| l sync.Mutex | ||
| nextIdx int | ||
| healthyInstances map[int]struct{} | ||
|
|
||
| startOnce sync.Once | ||
| stopOnce sync.Once | ||
| stopCh chan struct{} | ||
|
|
||
| logger *logging.Logger | ||
| } | ||
|
|
||
| // Implements host.Runtime. | ||
| func (lb *lbRuntime) ID() common.Namespace { | ||
| return lb.id | ||
| } | ||
|
|
||
| // Implements host.Runtime. | ||
| func (lb *lbRuntime) GetInfo(ctx context.Context) (*protocol.RuntimeInfoResponse, error) { | ||
| return lb.instances[0].GetInfo(ctx) | ||
| } | ||
|
|
||
| // Implements host.Runtime. | ||
| func (lb *lbRuntime) GetCapabilityTEE() (*node.CapabilityTEE, error) { | ||
| // TODO: This won't work when registration of all client runtimes is required. | ||
| return lb.instances[0].GetCapabilityTEE() | ||
| } | ||
|
|
||
| // shouldPropagateToAll checks whether the given runtime request should be propagated to all | ||
| // instances. | ||
| func shouldPropagateToAll(body *protocol.Body) bool { | ||
| switch { | ||
| case body.RuntimeConsensusSyncRequest != nil: | ||
| // Consensus view of all instances should be up to date as otherwise signed attestations | ||
| // will be stale, resulting in them being rejected. | ||
| return true | ||
| case body.RuntimeKeyManagerPolicyUpdateRequest != nil, | ||
| body.RuntimeKeyManagerStatusUpdateRequest != nil, | ||
| body.RuntimeKeyManagerQuotePolicyUpdateRequest != nil: | ||
| // Key manager updates should be propagated. | ||
| return true | ||
| default: | ||
| return false | ||
| } | ||
| } | ||
|
|
||
| // Implements host.Runtime. | ||
| func (lb *lbRuntime) Call(ctx context.Context, body *protocol.Body) (*protocol.Body, error) { | ||
| switch { | ||
| case shouldPropagateToAll(body): | ||
| // Propagate call to all instances. | ||
| type result struct { | ||
| rsp *protocol.Body | ||
| err error | ||
| } | ||
| resCh := make(chan *result) | ||
| for _, rt := range lb.instances { | ||
| rt := rt // Make sure goroutine below operates on the right instance. | ||
|
|
||
| go func() { | ||
| rsp, err := rt.Call(ctx, body) | ||
| resCh <- &result{ | ||
| rsp: rsp, | ||
| err: err, | ||
| } | ||
| }() | ||
| } | ||
|
|
||
| var ( | ||
| anyErr error | ||
| rsp *protocol.Body | ||
| ) | ||
| for range lb.instances { | ||
| res := <-resCh | ||
| // Return the response of the instance that finished last. Note that currently all of | ||
| // the propagated methods return a `protocol.Empty` response so this does not matter. | ||
| rsp = res.rsp | ||
| anyErr = errors.Join(anyErr, res.err) | ||
| } | ||
| if anyErr != nil { | ||
| return nil, anyErr | ||
| } | ||
| return rsp, nil | ||
| case body.RuntimeQueryRequest != nil, body.RuntimeCheckTxBatchRequest != nil: | ||
| // Load-balance queries. | ||
| idx, err := lb.selectInstance() | ||
| if err != nil { | ||
| return nil, err | ||
| } | ||
|
|
||
| lbRequestCount.With(prometheus.Labels{ | ||
| "runtime": lb.id.String(), | ||
| "lb_instance": fmt.Sprintf("%d", idx), | ||
| }).Inc() | ||
|
|
||
| return lb.instances[idx].Call(ctx, body) | ||
| default: | ||
| // Propagate only to the first instance. | ||
| return lb.instances[0].Call(ctx, body) | ||
| } | ||
| } | ||
|
|
||
| func (lb *lbRuntime) selectInstance() (int, error) { | ||
| lb.l.Lock() | ||
| defer lb.l.Unlock() | ||
|
|
||
| for attempt := 0; attempt < len(lb.instances); attempt++ { | ||
| idx := lb.nextIdx | ||
| lb.nextIdx = (lb.nextIdx + 1) % len(lb.instances) | ||
|
|
||
| if _, healthy := lb.healthyInstances[idx]; healthy { | ||
| return idx, nil | ||
| } | ||
| } | ||
|
|
||
| return 0, fmt.Errorf("host/loadbalance: no healthy instances available") | ||
| } | ||
|
|
||
| // Implements host.Runtime. | ||
| func (lb *lbRuntime) UpdateCapabilityTEE() { | ||
| for _, rt := range lb.instances { | ||
| rt.UpdateCapabilityTEE() | ||
| } | ||
| } | ||
|
|
||
| // Implements host.Runtime. | ||
| func (lb *lbRuntime) WatchEvents() (<-chan *host.Event, pubsub.ClosableSubscription) { | ||
| return lb.instances[0].WatchEvents() | ||
| } | ||
|
|
||
| // Implements host.Runtime. | ||
| func (lb *lbRuntime) Start() { | ||
| lb.startOnce.Do(func() { | ||
| for idx, rt := range lb.instances { | ||
| idx := idx | ||
| rt := rt // Make sure goroutine below operates on the right instance. | ||
|
|
||
| // Subscribe to runtime events before starting runtime to make sure we don't miss the | ||
| // started event. | ||
| evCh, sub := rt.WatchEvents() | ||
|
|
||
| // Start a goroutine to monitor whether an instance is healthy. | ||
| go func() { | ||
| defer sub.Close() | ||
|
|
||
| for { | ||
| select { | ||
| case ev := <-evCh: | ||
| switch { | ||
| case ev.Started != nil: | ||
| // Mark instance as available. | ||
| lb.logger.Info("instance is available", | ||
| "instance", idx, | ||
| ) | ||
|
|
||
| lb.l.Lock() | ||
| lb.healthyInstances[idx] = struct{}{} | ||
| lb.l.Unlock() | ||
| case ev.FailedToStart != nil, ev.Stopped != nil: | ||
| // Mark instance as failed. | ||
| lb.logger.Warn("instance is no longer available", | ||
| "instance", idx, | ||
| ) | ||
|
|
||
| lb.l.Lock() | ||
| delete(lb.healthyInstances, idx) | ||
| lb.l.Unlock() | ||
| default: | ||
| } | ||
|
|
||
| // Update healthy instance count metrics. | ||
| lb.l.Lock() | ||
| healthyInstanceCount := len(lb.healthyInstances) | ||
| lb.l.Unlock() | ||
|
|
||
| lbHealthyInstanceCount.With(prometheus.Labels{ | ||
| "runtime": lb.id.String(), | ||
| }).Set(float64(healthyInstanceCount)) | ||
| case <-lb.stopCh: | ||
| return | ||
| } | ||
| } | ||
| }() | ||
|
|
||
| rt.Start() | ||
| } | ||
| }) | ||
| } | ||
|
|
||
| // Implements host.Runtime. | ||
| func (lb *lbRuntime) Abort(ctx context.Context, force bool) error { | ||
| // We don't know which instance to abort, so we abort all instances. | ||
| errCh := make(chan error) | ||
| for _, rt := range lb.instances { | ||
| rt := rt // Make sure goroutine below operates on the right instance. | ||
|
|
||
| go func() { | ||
| errCh <- rt.Abort(ctx, force) | ||
| }() | ||
| } | ||
|
|
||
| var anyErr error | ||
| for range lb.instances { | ||
| err := <-errCh | ||
| anyErr = errors.Join(anyErr, err) | ||
| } | ||
| return anyErr | ||
| } | ||
|
|
||
| // Implements host.Runtime. | ||
| func (lb *lbRuntime) Stop() { | ||
| lb.stopOnce.Do(func() { | ||
| close(lb.stopCh) | ||
|
|
||
| for _, rt := range lb.instances { | ||
| rt.Stop() | ||
| } | ||
| }) | ||
| } | ||
|
|
||
| type lbProvisioner struct { | ||
| inner host.Provisioner | ||
| cfg Config | ||
| } | ||
|
|
||
| // Implements host.Provisioner. | ||
| func (lb *lbProvisioner) NewRuntime(cfg host.Config) (host.Runtime, error) { | ||
| if lb.cfg.NumInstances < 2 { | ||
| // This should never happen as the provisioner constructor made sure, but just to be safe. | ||
| return nil, fmt.Errorf("host/loadbalance: number of instances must be at least two") | ||
| } | ||
|
|
||
| // Use the inner provisioner to provision multiple runtimes. | ||
| var instances []host.Runtime | ||
| for i := 0; i < lb.cfg.NumInstances; i++ { | ||
| rt, err := lb.inner.NewRuntime(cfg) | ||
| if err != nil { | ||
| return nil, fmt.Errorf("host/loadbalance: failed to provision instance %d: %w", i, err) | ||
| } | ||
|
|
||
| instances = append(instances, rt) | ||
| } | ||
|
|
||
| return &lbRuntime{ | ||
| id: cfg.Bundle.Manifest.ID, | ||
| instances: instances, | ||
| healthyInstances: make(map[int]struct{}), | ||
| stopCh: make(chan struct{}), | ||
| logger: logging.GetLogger("runtime/host/loadbalance").With("runtime_id", cfg.Bundle.Manifest.ID), | ||
| }, nil | ||
| } | ||
|
|
||
| // Implements host.Provisioner. | ||
| func (lb *lbProvisioner) Name() string { | ||
| return fmt.Sprintf("load-balancer[%d]/%s", lb.cfg.NumInstances, lb.inner.Name()) | ||
| } | ||
|
|
||
| // New creates a load-balancing runtime provisioner. | ||
| func New(inner host.Provisioner, cfg Config) host.Provisioner { | ||
| if cfg.NumInstances < 2 { | ||
| // If there is only a single instance configured just return the inner provisioner. | ||
| return inner | ||
| } | ||
|
|
||
| initMetrics() | ||
|
|
||
| return &lbProvisioner{ | ||
| inner: inner, | ||
| cfg: cfg, | ||
| } | ||
| } |
Oops, something went wrong.