feat(sandbox): pre-pause guest reclaim via envd

packages/orchestrator/cmd/resume-build/main.go

@@ -72,8 +72,18 @@ func main() {
 	cmdSignalPause := flag.String("cmd-signal-pause", "", "execute command in sandbox, then wait for SIGUSR1 before pausing")
 	optimize := flag.Bool("optimize", false, "collect fresh prefetch mapping after pause (resumes snapshot to record page faults)")
 
+	// Enables the pre-pause reclaim chain with sensible per-step caps.
+	reclaim := flag.Bool("reclaim", false, "enable pre-pause reclaim chain (sync 500ms, drop_caches 200ms, compact 1s, fstrim 500ms)")
+
 	flag.Parse()
 
+	if *reclaim {
+		featureflags.NewIntFlag("reclaim-sync-timeout-ms", 500)
+		featureflags.NewIntFlag("reclaim-drop-caches-timeout-ms", 200)
+		featureflags.NewIntFlag("reclaim-compact-memory-timeout-ms", 1000)
+		featureflags.NewIntFlag("reclaim-fstrim-timeout-ms", 500)
+	}
+
 	if *fromBuild == "" {
 		log.Fatal("-from-build required")
 	}

packages/orchestrator/pkg/sandbox/reclaim.go

@@ -0,0 +1,116 @@
+package sandbox
+
+import (
+	"context"
+	"fmt"
+	"net/http"
+	"strconv"
+	"strings"
+	"time"
+
+	"connectrpc.com/connect"
+	"go.uber.org/zap"
+
+	"github.com/e2b-dev/infra/packages/shared/pkg/consts"
+	"github.com/e2b-dev/infra/packages/shared/pkg/featureflags"
+	"github.com/e2b-dev/infra/packages/shared/pkg/grpc"
+	"github.com/e2b-dev/infra/packages/shared/pkg/grpc/envd/process"
+	"github.com/e2b-dev/infra/packages/shared/pkg/grpc/envd/process/processconnect"
+	"github.com/e2b-dev/infra/packages/shared/pkg/logger"
+)
+
+type reclaimStep struct {
+	flag featureflags.IntFlag
+	cmd  string
+}
+
+// Order matters: sync makes drop_caches more effective; drop_caches gives
+// compact_memory more headroom; fstrim wants a stable FS view.
+var reclaimSteps = []reclaimStep{
+	{featureflags.ReclaimSyncTimeoutMs, "sync"},
+	{featureflags.ReclaimDropCachesTimeoutMs, "echo 3 > /proc/sys/vm/drop_caches"},
+	{featureflags.ReclaimCompactMemoryTimeoutMs, "echo 1 > /proc/sys/vm/compact_memory"},
+	{featureflags.ReclaimFstrimTimeoutMs, "fstrim -av"},
+}
+
+// Slack added to the sum of per-step caps to absorb shell start /
+// envd round-trip overhead.
+const reclaimOuterSlack = 500 * time.Millisecond
+
+// buildReclaimScript composes a chain where each step has its own
+// `timeout -s KILL` ceiling. Steps with cap=0 are skipped. Returns
+// ("", 0) when every step is disabled.
+func (s *Sandbox) buildReclaimScript(ctx context.Context) (string, time.Duration) {
+	var (
+		parts []string
+		sum   time.Duration
+	)
+	for _, st := range reclaimSteps {
+		ms := s.featureFlags.IntFlag(ctx, st.flag)
+		if ms <= 0 {
+			continue
+		}
+		// `timeout` accepts fractional seconds (s/m/h/d), not `ms`. Output
+		// is dropped; non-zero status is captured into `rc` so the final
+		// exit code surfaces failures without short-circuiting later steps.
+		secs := float64(ms) / 1000.0
+		parts = append(parts, fmt.Sprintf("timeout -s KILL %.3f sh -c %q >/dev/null 2>&1 || rc=$?", secs, st.cmd))
+		sum += time.Duration(ms) * time.Millisecond
+	}
+	if len(parts) == 0 {
+		return "", 0
+	}
+
+	return "rc=0; " + strings.Join(parts, "; ") + "; exit $rc", sum + reclaimOuterSlack
+}
+
+// bestEffortReclaim asks envd to reclaim guest memory + disk before pause.
+// Per-step output is silenced inside the guest; we only log when envd
+// itself errors or the script reports a non-zero exit code.
+func (s *Sandbox) bestEffortReclaim(ctx context.Context) {
+	script, timeout := s.buildReclaimScript(ctx)
+	if script == "" {
+		return
+	}
+
+	ctx, span := tracer.Start(ctx, "envd-reclaim")
+	defer span.End()
+
+	rcCtx, cancel := context.WithTimeout(ctx, timeout)
+	defer cancel()
+
+	addr := fmt.Sprintf("http://%s:%d", s.Slot.HostIPString(), consts.DefaultEnvdServerPort)
+	pc := processconnect.NewProcessClient(&http.Client{Transport: sandboxHttpClient.Transport}, addr)
+
+	req := connect.NewRequest(&process.StartRequest{
+		Process: &process.ProcessConfig{Cmd: "/bin/bash", Args: []string{"-c", script}},
+	})
+	req.Header().Set("Connect-Timeout-Ms", strconv.FormatInt(int64(timeout/time.Millisecond), 10))
+	if s.Config.Envd.AccessToken != nil {
+		req.Header().Set("X-Access-Token", *s.Config.Envd.AccessToken)
+	}
+	grpc.SetUserHeader(req.Header(), "root")
+
+	stream, err := pc.Start(rcCtx, req)
+	if err != nil {
+		logger.L().Warn(ctx, "envd reclaim failed", logger.WithSandboxID(s.Runtime.SandboxID), zap.Error(err))
+
+		return
+	}
+	defer stream.Close()
+
+	var exitCode int32
+	for stream.Receive() {
+		if end := stream.Msg().GetEvent().GetEnd(); end != nil {
+			exitCode = end.GetExitCode()
+		}
+	}
+	if err := stream.Err(); err != nil {
+		logger.L().Warn(ctx, "envd reclaim stream error", logger.WithSandboxID(s.Runtime.SandboxID), zap.Error(err))
+
+		return
+	}
+	if exitCode != 0 {
+		logger.L().Warn(ctx, "envd reclaim non-zero exit", logger.WithSandboxID(s.Runtime.SandboxID), zap.Int32("exit_code", exitCode))
+	}
+}

packages/orchestrator/pkg/sandbox/sandbox.go

@@ -217,6 +217,8 @@ type Sandbox struct {
 	files   *storage.SandboxFiles
 	cleanup *Cleanup
 
+	featureFlags *featureflags.Client
+
 	process      *fc.Process
 	cgroupHandle *cgroup.CgroupHandle
 
@@ -457,7 +459,8 @@ func (f *Factory) CreateSandbox(
 		files:    sandboxFiles,
 		process:  fcHandle,
 
-		cleanup: cleanup,
+		cleanup:      cleanup,
+		featureFlags: f.featureFlags,
 
 		APIStoredConfig: apiConfigToStore,
 
@@ -797,7 +800,8 @@ func (f *Factory) ResumeSandbox(
 		files:    sandboxFiles,
 		process:  fcHandle,
 
-		cleanup: cleanup,
+		cleanup:      cleanup,
+		featureFlags: f.featureFlags,
 
 		APIStoredConfig: apiConfigToStore,
 		CABundle:        f.egressProxy.CABundle(),
@@ -1051,6 +1055,11 @@ func (s *Sandbox) Pause(
 	// Stop the health check before pausing the VM
 	s.Checks.Stop()
 
+	// Best-effort pre-pause guest reclaim (sync, drop_caches, compact_memory,
+	// fstrim) on the live VM via envd. Per-step caps are LD-flag-driven; all
+	// default to 0 which disables the chain entirely. Non-fatal.
+	s.bestEffortReclaim(ctx)
+
 	if err := s.process.Pause(ctx); err != nil {
 		return nil, fmt.Errorf("failed to pause VM: %w", err)
 	}

packages/shared/pkg/featureflags/flags.go

@@ -157,7 +157,16 @@ var (
 	BestOfKMaxOvercommit          = NewIntFlag("best-of-k-max-overcommit", 400)              // Default R=4 (stored as percentage, max over-commit ratio)
 	BestOfKAlpha                  = NewIntFlag("best-of-k-alpha", 50)                        // Default Alpha=0.5 (stored as percentage for int flag, current usage weight)
 	EnvdInitTimeoutMilliseconds   = NewIntFlag("envd-init-request-timeout-milliseconds", 50) // Timeout for envd init request in milliseconds
-	HostStatsSamplingInterval     = NewIntFlag("host-stats-sampling-interval", 5000)         // Host stats sampling interval in milliseconds (default 5s)
+	// Per-step ceilings for the pre-pause guest reclaim chain (run via envd
+	// before snapshot). Each step is wrapped in `timeout -s KILL`; 0 skips it.
+	// All default to 0, so the feature is disabled until an operator opts in
+	// per step. A stuck step cannot starve the rest, so compact_memory always
+	// runs as long as its own cap is > 0.
+	ReclaimSyncTimeoutMs          = NewIntFlag("reclaim-sync-timeout-ms", 0)
+	ReclaimDropCachesTimeoutMs    = NewIntFlag("reclaim-drop-caches-timeout-ms", 0)
+	ReclaimCompactMemoryTimeoutMs = NewIntFlag("reclaim-compact-memory-timeout-ms", 0)
+	ReclaimFstrimTimeoutMs        = NewIntFlag("reclaim-fstrim-timeout-ms", 0)
+	HostStatsSamplingInterval     = NewIntFlag("host-stats-sampling-interval", 5000) // Host stats sampling interval in milliseconds (default 5s)
 	MaxCacheWriterConcurrencyFlag = NewIntFlag("max-cache-writer-concurrency", 10)
 
 	// BuildCacheMaxUsagePercentage the maximum percentage of the cache disk storage