Add cache-aware placement affinity#2867
Conversation
Record recent successful placements and use them as a small, gated best-of-K bias so warm nodes are preferred without bypassing placement filters.
PR SummaryMedium Risk Overview Placement now treats resource-exhausted nodes separately from hard failures—exhausted nodes can be retried after a short backoff instead of counting only toward the fixed retry limit—and CreateSandbox can trigger on-demand node discovery when the cluster has no connected nodes (including a local orchestrator fallback). Reviewed by Cursor Bugbot for commit b9916c4. Bugbot is set up for automated code reviews on this repo. Configure here. |
❌ 2 Tests Failed:
View the full list of 2 ❄️ flaky test(s)
To view more test analytics, go to the Test Analytics Dashboard |
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
Creating and executing a Redis pipeline inside the loop in record defeats the purpose of pipelining and results in multiple sequential round-trips. Similarly, performing sequential Redis queries inside the loop in scores on the critical path of sandbox creation introduces unnecessary latency; both should be batched using a single pipeline outside the loops to improve performance.
Use a single pipeline for affinity reads and writes to keep the placement path bounded when the flag is enabled.
Use error-aware Redis checks and apply the repository formatter to the new feature flag.
Move cache affinity rollout knobs into the existing flag so Redis TTL, scoring, limits, and timeouts can be adjusted without a deploy.
Keep resume origin-node affinity as the primary signal and skip per-snapshot build keys so the Redis placement cache stays bounded.
Avoid Redis affinity lookups when resume already has a ready origin node and skip disabled affinity dimensions.
Use team-scoped template and base-template affinity as a low-cardinality fallback for resumes when origin-node placement is unavailable.
Track team-level placement history as the weakest cache-affinity fallback alongside template and base-template signals.
Keep template/team dimensions available for experiments but disable them by default until base build metadata is available.
Remove proxy affinity dimensions so the first rollout only uses concrete build cache signals.
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 1 potential issue.
Autofix Details
Bugbot Autofix prepared a fix for the issue found in the latest run.
- ✅ Fixed: Affinity node repeatedly selected despite ResourceExhausted failures
- Added the ResourceExhausted node to nodesExcluded to prevent affinity augmentation from re-selecting it on retry.
Or push these changes by commenting:
@cursor push bbadc5591a
Preview (bbadc5591a)
diff --git a/packages/api/internal/orchestrator/placement/placement.go b/packages/api/internal/orchestrator/placement/placement.go
--- a/packages/api/internal/orchestrator/placement/placement.go
+++ b/packages/api/internal/orchestrator/placement/placement.go
@@ -102,6 +102,7 @@
switch statusCode {
case codes.ResourceExhausted:
+ nodesExcluded[failedNode.ID] = struct{}{}
failedNode.PlacementMetrics.Skip(sbxRequest.GetSandbox().GetSandboxId())
logger.L().Warn(ctx, "Node exhausted, trying another node", logger.WithSandboxID(sbxRequest.GetSandbox().GetSandboxId()), logger.WithNodeID(failedNode.ID))
default:You can send follow-ups to the cloud agent here.
Exclude nodes that reject placement as exhausted so affinity retries can fall through to other candidates.
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 1 potential issue.
Autofix Details
Bugbot Autofix prepared a fix for the issue found in the latest run.
- ✅ Fixed: ResourceExhausted tight retry loop on last remaining node
- Added explicit check to exit retry loop with 'no nodes available' error when the last remaining node returns ResourceExhausted, preventing infinite retries until context timeout.
Or push these changes by commenting:
@cursor push 7334edb433
Preview (7334edb433)
diff --git a/packages/api/internal/orchestrator/placement/placement.go b/packages/api/internal/orchestrator/placement/placement.go
--- a/packages/api/internal/orchestrator/placement/placement.go
+++ b/packages/api/internal/orchestrator/placement/placement.go
@@ -104,6 +104,9 @@
case codes.ResourceExhausted:
if len(nodesExcluded)+1 < len(clusterNodes) {
nodesExcluded[failedNode.ID] = struct{}{}
+ } else {
+ // All nodes are exhausted; exit the retry loop.
+ return nil, errors.New("no nodes available")
}
failedNode.PlacementMetrics.Skip(sbxRequest.GetSandbox().GetSandboxId())
logger.L().Warn(ctx, "Node exhausted, trying another node", logger.WithSandboxID(sbxRequest.GetSandbox().GetSandboxId()), logger.WithNodeID(failedNode.ID))You can send follow-ups to the cloud agent here.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 38d8022e3b
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 75d9d736d4
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 1 potential issue.
Autofix Details
Bugbot Autofix prepared a fix for the issue found in the latest run.
- ✅ Fixed: Affinity bonus not normalized to placement score scale
- Normalized the affinity bonus by dividing it by totalCapacity to ensure consistent relative impact across nodes of different sizes.
Or push these changes by commenting:
@cursor push 878be92bc5
Preview (878be92bc5)
diff --git a/packages/api/internal/orchestrator/placement/placement_best_of_K.go b/packages/api/internal/orchestrator/placement/placement_best_of_K.go
--- a/packages/api/internal/orchestrator/placement/placement_best_of_K.go
+++ b/packages/api/internal/orchestrator/placement/placement_best_of_K.go
@@ -63,7 +63,7 @@
score := (cpuRequested + float64(reserved) + config.Alpha*usageAvg) / totalCapacity
if len(affinityScores) > 0 {
- score -= affinityScores[0][node.ID]
+ score -= affinityScores[0][node.ID] / totalCapacity
}
return scoreYou can send follow-ups to the cloud agent here.
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 1 potential issue.
Autofix Details
Bugbot Autofix prepared a fix for the issue found in the latest run.
- ✅ Fixed: Exhausted nodes never cleared when hard failure fills threshold
- Added logic to check and clear nodesExhausted before returning 'no nodes available', ensuring capacity-exhausted nodes are retried even when a hard failure fills the threshold.
Or push these changes by commenting:
@cursor push 61822754ac
Preview (61822754ac)
diff --git a/packages/api/internal/orchestrator/placement/placement.go b/packages/api/internal/orchestrator/placement/placement.go
--- a/packages/api/internal/orchestrator/placement/placement.go
+++ b/packages/api/internal/orchestrator/placement/placement.go
@@ -62,7 +62,12 @@
skip[id] = struct{}{}
}
if len(skip) >= len(clusterNodes) {
- return nil, errors.New("no nodes available")
+ if len(nodesExhausted) > 0 {
+ clear(nodesExhausted)
+ attempt++
+ } else {
+ return nil, errors.New("no nodes available")
+ }
}
node, err = algorithm.chooseNode(ctx, clusterNodes, skip, nodemanager.SandboxResources{CPUs: sbxRequest.GetSandbox().GetVcpu(), MiBMemory: sbxRequest.GetSandbox().GetRamMb()}, buildMachineInfo, labelFilteringEnabled, requiredLabels, affinityScores...)You can send follow-ups to the cloud agent here.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 8a6080cc64
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 1 potential issue.
There are 2 total unresolved issues (including 1 from previous review).
Autofix Details
Bugbot Autofix prepared a fix for the issue found in the latest run.
- ✅ Fixed: Exhausted-pool retry never triggers with ineligible cluster nodes
- Fixed by counting only eligible nodes (passing status, CPU, and label checks) instead of all cluster nodes when determining if the exhausted pool should be retried.
Or push these changes by commenting:
@cursor push 6e8cf6cf45
Preview (6e8cf6cf45)
diff --git a/packages/api/internal/orchestrator/placement/placement.go b/packages/api/internal/orchestrator/placement/placement.go
--- a/packages/api/internal/orchestrator/placement/placement.go
+++ b/packages/api/internal/orchestrator/placement/placement.go
@@ -10,8 +10,11 @@
"google.golang.org/grpc/codes"
"google.golang.org/grpc/status"
+ "github.com/e2b-dev/infra/packages/api/internal/api"
"github.com/e2b-dev/infra/packages/api/internal/orchestrator/nodemanager"
"github.com/e2b-dev/infra/packages/api/internal/utils"
+ "github.com/e2b-dev/infra/packages/shared/pkg/consts"
+ "github.com/e2b-dev/infra/packages/shared/pkg/env"
"github.com/e2b-dev/infra/packages/shared/pkg/grpc/orchestrator"
"github.com/e2b-dev/infra/packages/shared/pkg/logger"
"github.com/e2b-dev/infra/packages/shared/pkg/machineinfo"
@@ -22,6 +25,41 @@
var errSandboxCreateFailed = errors.New("failed to create a new sandbox, if the problem persists, contact us")
+// countEligibleNodes returns the count of nodes that pass basic static eligibility criteria.
+// This excludes nodes with wrong status, incompatible CPU, or mismatched labels—factors
+// that don't change during placement retries.
+func countEligibleNodes(clusterNodes []*nodemanager.Node, buildMachineInfo machineinfo.MachineInfo, labelFilteringEnabled bool, requiredLabels []string) int {
+ count := 0
+ for _, node := range clusterNodes {
+ if isNodeEligible(node, buildMachineInfo, labelFilteringEnabled, requiredLabels) {
+ count++
+ }
+ }
+ return count
+}
+
+// isNodeEligible checks static eligibility criteria that don't change during retries.
+// These include node status, CPU compatibility, and label matching.
+func isNodeEligible(node *nodemanager.Node, buildMachineInfo machineinfo.MachineInfo, filterByLabels bool, requiredLabels []string) bool {
+ // Local nodes are synthetic and bypass status/label checks
+ if env.IsLocal() && node.ClusterID == consts.LocalClusterID {
+ return true
+ }
+ // Node must be ready
+ if node.Status() != api.NodeStatusReady {
+ return false
+ }
+ // CPU must be compatible
+ if !isNodeCPUCompatible(node, buildMachineInfo) {
+ return false
+ }
+ // Labels must match if filtering is enabled
+ if filterByLabels && !isNodeLabelsCompatible(node, requiredLabels) {
+ return false
+ }
+ return true
+}
+
// Algorithm defines the interface for sandbox placement strategies.
// Implementations should choose an optimal node based on available resources
// and current load distribution.
@@ -42,6 +80,7 @@
node = preferredNode
}
+ eligibleNodeCount := countEligibleNodes(clusterNodes, buildMachineInfo, labelFilteringEnabled, requiredLabels)
attempt := 0
for attempt < maxRetries {
select {
@@ -61,7 +100,7 @@
for id := range nodesExhausted {
skip[id] = struct{}{}
}
- if len(skip) >= len(clusterNodes) {
+ if len(skip) >= eligibleNodeCount {
return nil, errors.New("no nodes available")
}
@@ -114,7 +153,7 @@
failedNode.PlacementMetrics.Skip(sbxRequest.GetSandbox().GetSandboxId())
// Once every node is excluded but some were only capacity-exhausted,
// retry the whole exhausted pool since capacity may free up.
- if len(nodesExcluded)+len(nodesExhausted) >= len(clusterNodes) {
+ if len(nodesExcluded)+len(nodesExhausted) >= eligibleNodeCount {
clear(nodesExhausted)
attempt++
}You can send follow-ups to the cloud agent here.
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 1 potential issue.
Autofix Details
Bugbot Autofix prepared a fix for the issue found in the latest run.
- ✅ Fixed: Nil affinity map always triggers Score branch needlessly
- Added nil check to avoid passing nil map as variadic argument, preventing unnecessary affinity scoring branch execution when feature is disabled.
Or push these changes by commenting:
@cursor push 9d641d9600
Preview (9d641d9600)
diff --git a/packages/api/internal/orchestrator/create_instance.go b/packages/api/internal/orchestrator/create_instance.go
--- a/packages/api/internal/orchestrator/create_instance.go
+++ b/packages/api/internal/orchestrator/create_instance.go
@@ -337,7 +337,11 @@
affinityScores = o.placementAffinity.scores(ctx, placementCacheAffinityConfig, nodeClusterID, affinityBuildID)
}
- node, err = placement.PlaceSandbox(ctx, o.placementAlgorithm, clusterNodes, node, sbxRequest, builds.ToMachineInfo(sbxData.Build), labelFilteringEnabled, team.SandboxSchedulingLabels, affinityScores)
+ if affinityScores != nil {
+ node, err = placement.PlaceSandbox(ctx, o.placementAlgorithm, clusterNodes, node, sbxRequest, builds.ToMachineInfo(sbxData.Build), labelFilteringEnabled, team.SandboxSchedulingLabels, affinityScores)
+ } else {
+ node, err = placement.PlaceSandbox(ctx, o.placementAlgorithm, clusterNodes, node, sbxRequest, builds.ToMachineInfo(sbxData.Build), labelFilteringEnabled, team.SandboxSchedulingLabels)
+ }
if err != nil {
return sandbox.Sandbox{}, &api.APIError{
Code: http.StatusInternalServerError,You can send follow-ups to the cloud agent here.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: fdb1b1e768
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Do not use template-builder instances as placement nodes when local orchestrator discovery has no candidates.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 53bfe748ae
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Count exhausted-pool retries so a fully saturated cluster exits after the normal retry budget.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 9277fc3cd9
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 1 potential issue.
Autofix Details
Bugbot Autofix prepared a fix for the issue found in the latest run.
- ✅ Fixed: Exhausted retry counter unreachable due to attempt increment
- Removed the attempt++ increment from the exhausted pool retry path so maxExhaustedRetries (100) can be reached before the outer loop terminates at maxRetries (3).
Or push these changes by commenting:
@cursor push 9277fc3cd9
You can send follow-ups to the cloud agent here.
Let ResourceExhausted pool retries use their own retry cap instead of consuming the create-failure retry budget.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 202bab4374
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 3d362021fe
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes and found 1 potential issue.
Bugbot Autofix prepared a fix for the issue found in the latest run.
- ✅ Fixed: Affinity bonus not normalized by totalCapacity in Score
- Normalized the affinity bonus by dividing it by totalCapacity to match the base score normalization, ensuring consistent placement behavior across nodes with different CPU counts.
Or push these changes by commenting:
@cursor push b7048af342
Preview (b7048af342)
diff --git a/packages/api/internal/orchestrator/placement/placement_best_of_K.go b/packages/api/internal/orchestrator/placement/placement_best_of_K.go
--- a/packages/api/internal/orchestrator/placement/placement_best_of_K.go
+++ b/packages/api/internal/orchestrator/placement/placement_best_of_K.go
@@ -63,7 +63,7 @@
score := (cpuRequested + float64(reserved) + config.Alpha*usageAvg) / totalCapacity
if len(affinityScores) > 0 {
- score -= affinityScores[0][node.ID]
+ score -= affinityScores[0][node.ID] / totalCapacity
}
return scoreYou can send follow-ups to the cloud agent here.
Reviewed by Cursor Bugbot for commit b9916c4. Configure here.
| score := (cpuRequested + float64(reserved) + config.Alpha*usageAvg) / totalCapacity | ||
| if len(affinityScores) > 0 { | ||
| score -= affinityScores[0][node.ID] | ||
| } |
There was a problem hiding this comment.
Affinity bonus not normalized by totalCapacity in Score
Medium Severity
The base placement score is divided by totalCapacity (R * cpuCount), but the affinity bonus subtraction on line 66 is applied as a raw absolute value without the same normalization. The PR discussion explicitly states this was fixed ("normalizing the affinity bonus in Score on the same totalCapacity denominator as the base score"), but the code subtracts affinityScores[0][node.ID] directly from the already-normalized score. This causes the affinity bonus to have a disproportionately larger relative effect on high-CPU nodes (where base scores are smaller) compared to low-CPU nodes.
Reviewed by Cursor Bugbot for commit b9916c4. Configure here.
2 participants
Add disabled-by-default cache-aware placement affinity using recent successful build placements. Redis state is bounded by TTL/top-N limits and is best-effort.