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

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

@@ -7,7 +7,6 @@ import (
 	"fmt"
 	"log"
 	"math"
-	"net/http"
 	"os"
 	"os/signal"
 	"path/filepath"
@@ -16,7 +15,6 @@ import (
 	"syscall"
 	"time"
 
-	"connectrpc.com/connect"
 	"github.com/containernetworking/plugins/pkg/ns"
 	"github.com/coreos/go-iptables/iptables"
 	"github.com/google/uuid"
@@ -39,11 +37,8 @@ import (
 	"github.com/e2b-dev/infra/packages/orchestrator/pkg/tcpfirewall"
 	"github.com/e2b-dev/infra/packages/orchestrator/pkg/template/build/core/rootfs"
 	"github.com/e2b-dev/infra/packages/orchestrator/pkg/template/metadata"
-	"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"
 	sbxlogger "github.com/e2b-dev/infra/packages/shared/pkg/logger/sandbox"
 	"github.com/e2b-dev/infra/packages/shared/pkg/storage"
@@ -73,8 +68,18 @@ func main() {
 	optimize := flag.Bool("optimize", false, "collect fresh prefetch mapping after pause (resumes snapshot to record page faults)")
 	shell := flag.Bool("shell", false, "attach an interactive PTY shell via envd (no sshd required in the sandbox)")
 
+	// 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.NewDurationFlag("reclaim-sync-timeout", 500*time.Millisecond)
+		featureflags.NewDurationFlag("reclaim-drop-caches-timeout", 200*time.Millisecond)
+		featureflags.NewDurationFlag("reclaim-compact-memory-timeout", time.Second)
+		featureflags.NewDurationFlag("reclaim-fstrim-timeout", 500*time.Millisecond)
+	}
+
 	if *fromBuild == "" {
 		log.Fatal("-from-build required")
 	}
@@ -1026,7 +1031,9 @@ func run(ctx context.Context, buildID string, iterations int, coldStart, noPrefe
 	if verbose {
 		logLevel = ldlog.Info
 	}
-	flags, _ := featureflags.NewClientWithLogLevel(logLevel)
+	// Always use the offline data source so per-run flag overrides (e.g. set
+	// from -reclaim) take effect regardless of LAUNCH_DARKLY_API_KEY.
+	flags, _ := featureflags.NewOfflineClient(logLevel)
 
 	sandboxes := sandbox.NewSandboxesMap()
 
@@ -1184,32 +1191,10 @@ func printTemplateInfo(ctx context.Context, tmpl template.Template, meta metadat
 	}
 }
 
-// runCommandInSandbox runs a command inside the sandbox via envd
+// runCommandInSandbox runs a command inside the sandbox via envd as a
+// login shell so /etc/profile is sourced.
 func runCommandInSandbox(ctx context.Context, sbx *sandbox.Sandbox, command string) error {
-	// Connect directly to envd on the sandbox
-	envdURL := fmt.Sprintf("http://%s:%d", sbx.Slot.HostIPString(), consts.DefaultEnvdServerPort)
-
-	hc := http.Client{
-		Timeout:   10 * time.Minute,
-		Transport: sandbox.SandboxHttpTransport,
-	}
-
-	processC := processconnect.NewProcessClient(&hc, envdURL)
-
-	req := connect.NewRequest(&process.StartRequest{
-		Process: &process.ProcessConfig{
-			Cmd:  "/bin/bash",
-			Args: []string{"-l", "-c", command},
-		},
-	})
-	grpc.SetUserHeader(req.Header(), "root")
-
-	// Set access token if available
-	if sbx.Config.Envd.AccessToken != nil {
-		req.Header().Set("X-Access-Token", *sbx.Config.Envd.AccessToken)
-	}
-
-	stream, err := processC.Start(ctx, req)
+	stream, err := sbx.StartEnvdShell(ctx, []string{"-l", "-c", command}, "root", 0)
 	if err != nil {
 		return fmt.Errorf("failed to start process: %w", err)
 	}

packages/orchestrator/pkg/sandbox/envd_process.go

@@ -0,0 +1,46 @@
+package sandbox
+
+import (
+	"context"
+	"fmt"
+	"net/http"
+	"strconv"
+	"time"
+
+	"connectrpc.com/connect"
+
+	"github.com/e2b-dev/infra/packages/shared/pkg/consts"
+	"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"
+)
+
+// StartEnvdShell opens a streaming Process.Start call against this
+// sandbox's envd, running `/bin/sh` with the given args as `user`.
+// `/bin/sh` is used (not bash) so minimal guests without bash still
+// work. Caller chooses login-shell vs. plain (e.g. []{"-l","-c",cmd}
+// vs. []{"-c",script}). When timeout > 0 it sets `Connect-Timeout-Ms`
+// so envd kills the process at the deadline. Auth/user headers are
+// wired from sandbox config. Caller owns the returned stream.
+func (s *Sandbox) StartEnvdShell(
+	ctx context.Context,
+	shellArgs []string,
+	user string,
+	timeout time.Duration,
+) (*connect.ServerStreamForClient[process.StartResponse], error) {
+	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/sh", Args: shellArgs},
+	})
+	if timeout > 0 {
+		req.Header().Set("Connect-Timeout-Ms", strconv.FormatInt(timeout.Milliseconds(), 10))
+	}
+	if s.Config.Envd.AccessToken != nil {
+		req.Header().Set("X-Access-Token", *s.Config.Envd.AccessToken)
+	}
+	grpc.SetUserHeader(req.Header(), user)
+
+	return pc.Start(ctx, req)
+}

packages/orchestrator/pkg/sandbox/reclaim.go

@@ -0,0 +1,96 @@
+package sandbox
+
+import (
+	"context"
+	"fmt"
+	"strings"
+	"time"
+
+	"go.uber.org/zap"
+
+	"github.com/e2b-dev/infra/packages/shared/pkg/featureflags"
+	"github.com/e2b-dev/infra/packages/shared/pkg/logger"
+)
+
+type reclaimStep struct {
+	flag featureflags.DurationFlag
+	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.ReclaimSyncTimeout, "sync"},
+	{featureflags.ReclaimDropCachesTimeout, "echo 3 > /proc/sys/vm/drop_caches"},
+	{featureflags.ReclaimCompactMemoryTimeout, "echo 1 > /proc/sys/vm/compact_memory"},
+	{featureflags.ReclaimFstrimTimeout, "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 {
+		d := s.featureFlags.DurationFlag(ctx, st.flag)
+		if d <= 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.
+		parts = append(parts, fmt.Sprintf("timeout -s KILL %.3f sh -c %q >/dev/null 2>&1 || rc=$?", d.Seconds(), st.cmd))
+		sum += d
+	}
+	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()
+
+	stream, err := s.StartEnvdShell(rcCtx, []string{"-c", script}, "root", timeout)
+	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/client.go

@@ -89,6 +89,25 @@ func NewClientWithLogLevel(logLevel ldlog.LogLevel) (*Client, error) {
 	return &Client{ld: ldClient}, nil
 }
 
+// NewOfflineClient creates a client that always reads from the in-process
+// offline data source, ignoring LAUNCH_DARKLY_API_KEY. Use this in CLI/dev
+// tools (e.g. resume-build) so flag overrides set via NewIntFlag/NewDurationFlag
+// at startup take effect deterministically.
+func NewOfflineClient(logLevel ldlog.LogLevel) (*Client, error) {
+	cfg := ldclient.Config{
+		Logging:          ldcomponents.Logging().MinLevel(logLevel),
+		DataSource:       launchDarklyOfflineStore,
+		Events:           ldcomponents.NoEvents(),
+		DiagnosticOptOut: true,
+	}
+	ldClient, err := ldclient.MakeCustomClient("", cfg, 0)
+	if err != nil {
+		return nil, err
+	}
+
+	return &Client{ld: ldClient}, nil
+}
+
 func (c *Client) SetDeploymentName(deploymentName string) {
 	c.deploymentName = deploymentName
 }
@@ -113,6 +132,36 @@ func (c *Client) StringFlag(ctx context.Context, flag StringFlag, contexts ...ld
 	return getFlag(ctx, c.ld, c.ld.StringVariationCtx, flag, c.allContexts(contexts))
 }
 
+// DurationFlag reads a duration-valued flag. The underlying flag is stored as a
+// string (e.g. "500ms") and parsed via time.ParseDuration. Empty/invalid values
+// fall back to the configured default.
+func (c *Client) DurationFlag(ctx context.Context, flag DurationFlag, contexts ...ldcontext.Context) time.Duration {
+	if c.ld == nil {
+		logger.L().Info(ctx, "LaunchDarkly client is not initialized, returning fallback")
+
+		return flag.Fallback()
+	}
+
+	raw, err := c.ld.StringVariationCtx(ctx, flag.Key(), mergeContexts(ctx, c.allContexts(contexts)), "")
+	if err != nil {
+		logger.L().Warn(ctx, "error evaluating flag", zap.Error(err), zap.String("flag", flag.Key()))
+
+		return flag.Fallback()
+	}
+	if raw == "" {
+		return flag.Fallback()
+	}
+	d, err := time.ParseDuration(raw)
+	if err != nil {
+		logger.L().Warn(ctx, "invalid duration flag value, using fallback",
+			zap.String("flag", flag.Key()), zap.String("value", raw), zap.Error(err))
+
+		return flag.Fallback()
+	}
+
+	return d
+}
+
 type typedFlag[T any] interface {
 	Key() string
 	Fallback() T