diff --git a/etc/rc1 b/etc/rc1
index ab0d8811bffe..56474eb5f00b 100755
--- a/etc/rc1
+++ b/etc/rc1
@@ -54,7 +54,8 @@ if test $RANK -eq 0 -a "${FLUX_SCHED_MODULE}" != "none" \
     flux module load ${FLUX_SCHED_MODULE:-sched-simple}
 fi
 
-test $RANK -ne 0 || flux admin cleanup-push <<-EOT
+test $RANK -ne 0 -o "${FLUX_INSTANCE_RESTART}" = "t" \
+     || flux admin cleanup-push <<-EOT
 	flux queue stop --quiet
 	flux job cancelall --user=all --quiet -f --states RUN
 	flux queue idle --quiet
diff --git a/src/common/libkvs/kvs_getroot.h b/src/common/libkvs/kvs_getroot.h
index 5cdfe995e88c..94c370ba40db 100644
--- a/src/common/libkvs/kvs_getroot.h
+++ b/src/common/libkvs/kvs_getroot.h
@@ -21,7 +21,7 @@ flux_future_t *flux_kvs_getroot (flux_t *h, const char *ns, int flags);
 /* Decode KVS root hash response.
  *
  * treeobj - get the hash as an RFC 11 "dirref" object.
- * blobref - get the raw hash as a n RFC 10 "blobref".
+ * blobref - get the raw hash as an RFC 10 "blobref".
  * sequence - get the commit sequence number
  * owner - get the userid of the namespace owner
  */
diff --git a/src/modules/job-exec/Makefile.am b/src/modules/job-exec/Makefile.am
index 791262832a6c..618dcf47daf8 100644
--- a/src/modules/job-exec/Makefile.am
+++ b/src/modules/job-exec/Makefile.am
@@ -27,6 +27,8 @@ libbulk_exec_la_SOURCES = \
 job_exec_la_SOURCES = \
 	job-exec.h \
 	job-exec.c \
+	checkpoint.h \
+	checkpoint.c \
 	rset.c \
 	rset.h \
 	testexec.c \
diff --git a/src/modules/job-exec/checkpoint.c b/src/modules/job-exec/checkpoint.c
new file mode 100644
index 000000000000..055cb64396e8
--- /dev/null
+++ b/src/modules/job-exec/checkpoint.c
@@ -0,0 +1,253 @@
+/************************************************************\
+ * Copyright 2021 Lawrence Livermore National Security, LLC
+ * (c.f. AUTHORS, NOTICE.LLNS, COPYING)
+ *
+ * This file is part of the Flux resource manager framework.
+ * For details, see https://github.com/flux-framework.
+ *
+ * SPDX-License-Identifier: LGPL-3.0
+\************************************************************/
+
+/* Prototype checkpoint of running jobs KVS root refs
+ *
+ * DESCRIPTION
+ *
+ * Handle checkpoint of running job's guest KVS namescapes into the
+ * primary KVS.  This will allow the namespaces to be recreated if
+ * a job manager is brought down than back up.
+ *
+ * Also provide helper functions to get rootrefs from the checkpointed
+ * object.
+ *
+ * OPERATION
+ *
+ * Get the KVS rootrefs for all running jobs and commit to
+ * "job-exec.kvs-namespaces".
+ *
+ */
+
+#if HAVE_CONFIG_H
+#include "config.h"
+#endif
+#include <assert.h>
+#include <unistd.h>
+#include <errno.h>
+#include <flux/core.h>
+
+#include "src/common/libczmqcontainers/czmq_containers.h"
+
+#include "job-exec.h"
+#include "checkpoint.h"
+
+flux_future_t *checkpoint_get_rootrefs (flux_t *h)
+{
+    if (!h) {
+        errno = EINVAL;
+        return NULL;
+    }
+
+    return flux_kvs_lookup (h,
+                            NULL,
+                            0,
+                            "job-exec.kvs-namespaces");
+}
+
+char *checkpoint_find_rootref (flux_future_t *f,
+                               flux_jobid_t id,
+                               uint32_t owner)
+{
+    int saved_errno;
+    flux_t *h = flux_future_get_flux (f);
+    char *rv = NULL;
+    const char *rootrefs;
+    json_error_t error;
+    json_t *o = NULL;
+    size_t index;
+    json_t *value;
+
+    if (flux_kvs_lookup_get (f, &rootrefs) < 0) {
+        flux_log_error (h, "checkpoint_get_rootref: flux_kvs_lookup_get");
+        goto cleanup;
+    }
+
+    if (!(o = json_loads (rootrefs, 0, &error))) {
+        flux_log (h, LOG_ERR, "json_loads rootrefs: %s", error.text);
+        goto cleanup;
+    }
+
+    json_array_foreach (o, index, value) {
+        flux_jobid_t l_id;
+        uint32_t l_owner;
+        const char *rootref;
+
+        if (json_unpack (value,
+                         "{s:I s:i s:s}",
+                         "id", &l_id,
+                         "owner", &l_owner,
+                         "kvsroot", &rootref) < 0) {
+            flux_log (h, LOG_ERR, "json_unpack rootref: %s", error.text);
+            goto cleanup;
+        }
+        if (l_id == id && l_owner == owner) {
+            if (!(rv = strdup (rootref)))
+                goto cleanup;
+            break;
+        }
+    }
+
+cleanup:
+    saved_errno = errno;
+    json_decref (o);
+    errno = saved_errno;
+    return rv;
+}
+
+static int lookup_nsroots (flux_t *h, zhashx_t *jobs, flux_future_t **fp)
+{
+    struct jobinfo *job = zhashx_first (jobs);
+    flux_future_t *fall = NULL;
+    flux_future_t *f = NULL;
+
+    while (job) {
+        if (job->running) {
+            if (!fall) {
+                if (!(fall = flux_future_wait_all_create ()))
+                    goto cleanup;
+                flux_future_set_flux (fall, h);
+            }
+            if (!(f = flux_kvs_getroot (h, job->ns, 0)))
+                goto cleanup;
+            if (flux_future_aux_set (f, "jobinfo", job, NULL) < 0)
+                goto cleanup;
+            if (flux_future_push (fall, job->ns, f) < 0)
+                goto cleanup;
+            f = NULL;
+        }
+        job = zhashx_next (jobs);
+    }
+
+    (*fp) = fall;
+    return 0;
+
+cleanup:
+    flux_future_destroy (f);
+    flux_future_destroy (fall);
+    return -1;
+}
+
+static json_t *get_nsroots (flux_t *h, flux_future_t *fall)
+{
+    const char *child;
+    json_t *nsdata = NULL;
+    int saved_errno;
+
+    if (!(nsdata = json_array ())) {
+        errno = ENOMEM;
+        return NULL;
+    }
+
+    child = flux_future_first_child (fall);
+    while (child) {
+        flux_future_t *f = flux_future_get_child (fall, child);
+        struct jobinfo *job;
+        const char *blobref = NULL;
+        json_t *o = NULL;
+        if (!f)
+            goto cleanup;
+        if (!(job = flux_future_aux_get (f, "jobinfo")))
+            goto cleanup;
+        if (flux_kvs_getroot_get_blobref (f, &blobref) < 0)
+            goto cleanup;
+        if (!(o = json_pack ("{s:I s:i s:s}",
+                             "id", job->id,
+                             "owner", job->userid,
+                             "kvsroot", blobref))) {
+            errno = ENOMEM;
+            goto cleanup;
+        }
+        if (json_array_append (nsdata, o) < 0) {
+            json_decref (o);
+            errno = ENOMEM;
+            goto cleanup;
+        }
+        json_decref (o);
+        child = flux_future_next_child (fall);
+    }
+
+    return nsdata;
+cleanup:
+    saved_errno = errno;
+    json_decref (nsdata);
+    errno = saved_errno;
+    return NULL;
+}
+
+static int checkpoint_commit (flux_t *h, json_t *nsdata)
+{
+    flux_future_t *f = NULL;
+    flux_kvs_txn_t *txn = NULL;
+    char *s = NULL;
+    int rv = -1;
+
+    if (!(s = json_dumps (nsdata, JSON_COMPACT))) {
+        errno = ENOMEM;
+        goto cleanup;
+    }
+
+    if (!(txn = flux_kvs_txn_create ()))
+        goto cleanup;
+
+    if (flux_kvs_txn_put (txn,
+                          0,
+                          "job-exec.kvs-namespaces",
+                          s) < 0)
+        goto cleanup;
+
+    if (!(f = flux_kvs_commit (h, NULL, 0, txn)))
+        goto cleanup;
+
+    if (flux_future_get (f, NULL) < 0)
+        goto cleanup;
+
+    rv = 0;
+cleanup:
+    flux_future_destroy (f);
+    flux_kvs_txn_destroy (txn);
+    free (s);
+    return rv;
+}
+
+void checkpoint_running (flux_t *h, zhashx_t *jobs)
+{
+    flux_future_t *lookupf = NULL;
+    json_t *nsdata = NULL;
+
+    if (!h || !jobs)
+        return;
+
+    if (lookup_nsroots (h, jobs, &lookupf) < 0) {
+        flux_log_error (h, "failed to lookup ns root refs");
+        goto cleanup;
+    }
+
+    if (!lookupf)
+        return;
+
+    if (!(nsdata = get_nsroots (h, lookupf))) {
+        flux_log_error (h, "failure getting ns root refs");
+        goto cleanup;
+    }
+
+    if (checkpoint_commit (h, nsdata) < 0) {
+        flux_log_error (h, "failure committing ns checkpoint data");
+        goto cleanup;
+    }
+
+cleanup:
+    json_decref (nsdata);
+    flux_future_destroy (lookupf);
+}
+
+/*
+ * vi: tabstop=4 shiftwidth=4 expandtab
+ */
diff --git a/src/modules/job-exec/checkpoint.h b/src/modules/job-exec/checkpoint.h
new file mode 100644
index 000000000000..c4d036f5faac
--- /dev/null
+++ b/src/modules/job-exec/checkpoint.h
@@ -0,0 +1,30 @@
+/************************************************************\
+ * Copyright 2021 Lawrence Livermore National Security, LLC
+ * (c.f. AUTHORS, NOTICE.LLNS, COPYING)
+ *
+ * This file is part of the Flux resource manager framework.
+ * For details, see https://github.com/flux-framework.
+ *
+ * SPDX-License-Identifier: LGPL-3.0
+\************************************************************/
+
+#ifndef HAVE_JOB_EXEC_CHECKPOINT_H
+#define HAVE_JOB_EXEC_CHECKPOINT_H 1
+
+#include <flux/core.h>
+
+#include "src/common/libczmqcontainers/czmq_containers.h"
+#include "job-exec.h"
+
+flux_future_t *checkpoint_get_rootrefs (flux_t *h);
+
+char *checkpoint_find_rootref (flux_future_t *f,
+                               flux_jobid_t id,
+                               uint32_t owner);
+
+void checkpoint_running (flux_t *h, zhashx_t *jobs);
+
+#endif /* !HAVE_JOB_EXEC_CHECKPOINT_EXEC_H */
+
+/* vi: ts=4 sw=4 expandtab
+ */
diff --git a/src/modules/job-exec/exec.c b/src/modules/job-exec/exec.c
index 3bb837a9e749..eec88ad3e24f 100644
--- a/src/modules/job-exec/exec.c
+++ b/src/modules/job-exec/exec.c
@@ -111,7 +111,7 @@ static const char *job_get_cwd (struct jobinfo *job)
 static void start_cb (struct bulk_exec *exec, void *arg)
 {
     struct jobinfo *job = arg;
-    jobinfo_started (job, NULL);
+    jobinfo_started (job);
     /*  This is going to be really slow. However, it should at least
      *   work for now. We wait for all imp's to start, then send input
      */
diff --git a/src/modules/job-exec/job-exec.c b/src/modules/job-exec/job-exec.c
index 5e95cd3aeaec..7ece63eb4beb 100644
--- a/src/modules/job-exec/job-exec.c
+++ b/src/modules/job-exec/job-exec.c
@@ -26,7 +26,7 @@
  *
  * JOB INIT:
  *
- * On reciept of a start request, the exec service enters initialization
+ * On receipt of a start request, the exec service enters initialization
  * phase of the job, where the jobspec and R are fetched from the KVS,
  * and the guest namespace is created and linked from the primary
  * namespace. A guest.exec.eventlog is created with an initial "init"
@@ -99,6 +99,7 @@
 #include "src/common/libutil/errno_safe.h"
 
 #include "job-exec.h"
+#include "checkpoint.h"
 
 static double kill_timeout=5.0;
 
@@ -138,6 +139,7 @@ void jobinfo_decref (struct jobinfo *job)
         free (job->J);
         resource_set_destroy (job->R);
         json_decref (job->jobspec);
+        free (job->rootref);
         free (job);
         errno = saved_errno;
     }
@@ -290,7 +292,7 @@ static int jobinfo_send_release (struct jobinfo *job,
 }
 
 static int jobinfo_respond (flux_t *h, struct jobinfo *job,
-                            const char *event, int status)
+                            const char *event)
 {
     return flux_respond_pack (h, job->req, "{s:I s:s s:{}}",
                                            "id", job->id,
@@ -319,19 +321,26 @@ static void jobinfo_complete (struct jobinfo *job, const struct idset *ranks)
     }
 }
 
-void jobinfo_started (struct jobinfo *job, const char *fmt, ...)
+void jobinfo_started (struct jobinfo *job)
 {
     flux_t *h = job->ctx->h;
     if (h && job->req) {
-        va_list ap;
-        va_start (ap, fmt);
         job->running = 1;
-        va_end (ap);
-        if (jobinfo_respond (h, job, "start", 0) < 0)
+        if (jobinfo_respond (h, job, "start") < 0)
             flux_log_error (h, "jobinfo_started: flux_respond");
     }
 }
 
+void jobinfo_reattached (struct jobinfo *job)
+{
+    flux_t *h = job->ctx->h;
+    if (h && job->req) {
+        job->running = 1;
+        if (jobinfo_respond (h, job, "reattached") < 0)
+            flux_log_error (h, "jobinfo_reattach: flux_respond");
+    }
+}
+
 static void kill_timer_cb (flux_reactor_t *r, flux_watcher_t *w,
                            int revents, void *arg)
 {
@@ -594,7 +603,10 @@ static int jobinfo_finalize (struct jobinfo *job)
 
 static int jobinfo_start_execution (struct jobinfo *job)
 {
-    jobinfo_emit_event_pack_nowait (job, "starting", NULL);
+    if (job->reattach)
+        jobinfo_emit_event_pack_nowait (job, "re-starting", NULL);
+    else
+        jobinfo_emit_event_pack_nowait (job, "starting", NULL);
     /* Set started flag before calling 'start' method because we want to
      *  be sure to clean up properly if an exception occurs
      */
@@ -731,6 +743,7 @@ static void jobinfo_start_continue (flux_future_t *f, void *arg)
     }
     job->has_namespace = 1;
 
+
     /*  If an exception was received during startup, no need to continue
      *   with startup
      */
@@ -808,7 +821,9 @@ static void namespace_link (flux_future_t *fprev, void *arg)
         goto error;
     }
     flux_future_set_flux (cf, h);
-    if (!(f = jobinfo_emit_event_pack (job, "init", NULL))
+    if (!(f = jobinfo_emit_event_pack (job,
+                                       job->reattach ? "reattach" : "init",
+                                       NULL))
         || flux_future_push (cf, "emit event", f) < 0)
         goto error;
 
@@ -832,8 +847,16 @@ static flux_future_t *ns_create_and_link (flux_t *h,
     flux_future_t *f = NULL;
     flux_future_t *f2 = NULL;
 
-    if (!(f = flux_kvs_namespace_create (h, job->ns, job->userid, flags))
-        || !(f2 = flux_future_and_then (f, namespace_link, job))) {
+    if (job->reattach && job->rootref)
+        f = flux_kvs_namespace_create_with (h,
+                                            job->ns,
+                                            job->rootref,
+                                            job->userid,
+                                            flags);
+    else
+        f = flux_kvs_namespace_create (h, job->ns, job->userid, flags);
+
+    if (!f || !(f2 = flux_future_and_then (f, namespace_link, job))) {
         flux_log_error (h, "namespace_move: flux_future_and_then");
         flux_future_destroy (f);
         return NULL;
@@ -841,6 +864,54 @@ static flux_future_t *ns_create_and_link (flux_t *h,
     return f2;
 }
 
+static void get_rootref_cb (flux_future_t *fprev, void *arg)
+{
+    int saved_errno;
+    flux_t *h = flux_future_get_flux (fprev);
+    struct jobinfo *job = arg;
+    flux_future_t *f = NULL;
+
+    if (!(job->rootref = checkpoint_find_rootref (fprev,
+                                                  job->id,
+                                                  job->userid)))
+        flux_log (job->h,
+                  LOG_DEBUG,
+                  "checkpoint rootref not found: %ju",
+                  (uintmax_t)job->id);
+
+    /* if rootref not found, still create namespace */
+    if (!(f = ns_create_and_link (h, job, 0)))
+        goto error;
+
+    flux_future_continue (fprev, f);
+    flux_future_destroy (fprev);
+    return;
+error:
+    saved_errno = errno;
+    flux_future_destroy (f);
+    flux_future_continue_error (fprev, saved_errno, NULL);
+    flux_future_destroy (fprev);
+}
+
+static flux_future_t *ns_get_rootref (flux_t *h,
+                                      struct jobinfo *job,
+                                      int flags)
+{
+    flux_future_t *f = NULL;
+    flux_future_t *f2 = NULL;
+
+    if (!(f = checkpoint_get_rootrefs (h))) {
+        flux_log_error (h, "ns_get_rootref: checkpoint_get_rootrefs");
+        return NULL;
+    }
+    if (!f || !(f2 = flux_future_and_then (f, get_rootref_cb, job))) {
+        flux_log_error (h, "ns_get_rootref: flux_future_and_then");
+        flux_future_destroy (f);
+        return NULL;
+    }
+    return f2;
+}
+
 /*  Asynchronously fetch job data from KVS and create namespace.
  */
 static flux_future_t *jobinfo_start_init (struct jobinfo *job)
@@ -858,8 +929,12 @@ static flux_future_t *jobinfo_start_init (struct jobinfo *job)
         || flux_future_push (f, "J", f_kvs) < 0)) {
         goto err;
     }
-    if (!(f_kvs = ns_create_and_link (h, job, 0))
-        || flux_future_push (f, "ns", f_kvs))
+    if (job->reattach)
+        f_kvs = ns_get_rootref (h, job, 0);
+    else
+        f_kvs = ns_create_and_link (h, job, 0);
+
+    if (flux_future_push (f, "ns", f_kvs) < 0)
         goto err;
 
     return f;
@@ -916,10 +991,11 @@ static int job_start (struct job_exec_ctx *ctx, const flux_msg_t *msg)
 
     job->ctx = ctx;
 
-    if (flux_request_unpack (job->req, NULL, "{s:I s:i s:O}",
+    if (flux_request_unpack (job->req, NULL, "{s:I s:i s:O s:b}",
                                              "id", &job->id,
                                              "userid", &job->userid,
-                                             "jobspec", &job->jobspec) < 0) {
+                                             "jobspec", &job->jobspec,
+                                             "reattach", &job->reattach) < 0) {
         flux_log_error (ctx->h, "start: flux_request_unpack");
         jobinfo_decref (job);
         return -1;
@@ -1103,10 +1179,50 @@ static int configure_implementations (flux_t *h, int argc, char **argv)
     return 0;
 }
 
-static int unload_implementations (void)
+static int remove_running_ns (struct job_exec_ctx *ctx)
+{
+    struct jobinfo *job = zhashx_first (ctx->jobs);
+    flux_future_t *fall = NULL;
+    flux_future_t *f = NULL;
+    int rv = -1;
+
+    while (job) {
+        if (job->running) {
+            if (!fall) {
+                if (!(fall = flux_future_wait_all_create ()))
+                    goto cleanup;
+                flux_future_set_flux (fall, ctx->h);
+            }
+            if (!(f = flux_kvs_namespace_remove (ctx->h, job->ns)))
+                goto cleanup;
+            if (flux_future_push (fall, job->ns, f) < 0)
+                goto cleanup;
+            f = NULL;
+        }
+        job = zhashx_next (ctx->jobs);
+    }
+
+    if (fall) {
+        if (flux_future_wait_for (fall, -1.) < 0)
+            goto cleanup;
+    }
+
+    rv = 0;
+cleanup:
+    flux_future_destroy (f);
+    flux_future_destroy (fall);
+    return rv;
+}
+
+static int unload_implementations (struct job_exec_ctx *ctx)
 {
     struct exec_implementation *impl;
     int i = 0;
+    if (ctx && ctx->jobs) {
+        checkpoint_running (ctx->h, ctx->jobs);
+        if (remove_running_ns (ctx) < 0)
+            flux_log_error (ctx->h, "failed to remove guest namespaces");
+    }
     while ((impl = implementations[i]) && impl->name) {
         if (impl->unload)
              (*impl->unload) ();
@@ -1146,7 +1262,7 @@ int mod_main (flux_t *h, int argc, char **argv)
 
     rc = flux_reactor_run (flux_get_reactor (h), 0);
 out:
-    unload_implementations ();
+    unload_implementations (ctx);
 
     saved_errno = errno;
     if (flux_event_unsubscribe (h, "job-exception") < 0)
diff --git a/src/modules/job-exec/job-exec.h b/src/modules/job-exec/job-exec.h
index 2cbed57e9b66..3cf6a25a83a4 100644
--- a/src/modules/job-exec/job-exec.h
+++ b/src/modules/job-exec/job-exec.h
@@ -61,6 +61,7 @@ struct jobinfo {
     flux_t *              h;
     flux_jobid_t          id;
     char                  ns [64];   /* namespace string */
+    char                * rootref;   /* ns rootref if restart */
     const flux_msg_t *    req;       /* initial request */
     uint32_t              userid;    /* requesting userid */
     int                   flags;     /* job flags */
@@ -77,6 +78,7 @@ struct jobinfo {
     uint8_t               running:1;     /* all shells are running */
     uint8_t               finalizing:1;  /* in process of cleanup */
 
+    int                   reattach;      /* job-manager reattach attempt */
     int                   wait_status;
 
     struct eventlogger *  ev;           /* event batcher */
@@ -102,8 +104,11 @@ int jobinfo_emit_event_pack_nowait (struct jobinfo *job,
                                      const char *name,
                                      const char *fmt, ...);
 
-/* Emit  start event with optional note in jansson pack format */
-void jobinfo_started (struct jobinfo *job, const char *fmt, ...);
+/* Emit  start event */
+void jobinfo_started (struct jobinfo *job);
+
+/* Emit  reattached event */
+void jobinfo_reattached (struct jobinfo *job);
 
 /* Notify job-exec that ranks in idset `ranks` have completed
  *  with the given wait status
diff --git a/src/modules/job-exec/testexec.c b/src/modules/job-exec/testexec.c
index c8f6e7e71f2d..3033f527febd 100644
--- a/src/modules/job-exec/testexec.c
+++ b/src/modules/job-exec/testexec.c
@@ -32,6 +32,9 @@
  *                             event a testexec job. If job has unlimited
  *                             duration, then also wait for finish RPC or
  *                             job exception for job finish event.
+ *   "reattach_finish":i    - if reattached, assume job has finished
+ *                             and no longer has remaining time to run.
+ *                             Useful for testing job reattach.
  * }
  *
  */
@@ -40,8 +43,12 @@
 # include "config.h"
 #endif
 
+#include <time.h>
+#include <sys/time.h>
+
 #include "src/common/libutil/fsd.h"
 #include "src/common/libjob/job_hash.h"
+#include "src/common/libeventlog/eventlog.h"
 #include "src/common/libczmqcontainers/czmq_containers.h"
 #include "job-exec.h"
 
@@ -55,6 +62,7 @@ struct testexec_ctx {
 struct testconf {
     bool                  enabled;          /* test execution enabled       */
     int                   override;         /* wait for RPC for start event */
+    int                   reattach_finish;  /* if reattached, just finish job */
     double                run_duration;     /* duration of fake job in sec  */
     int                   wait_status;      /* reported status for "finish" */
     const char *          mock_exception;   /* fake excetion at this site   */
@@ -108,6 +116,7 @@ static int init_testconf (flux_t *h, struct testconf *conf, json_t *jobspec)
     /* get/set defaults */
     conf->run_duration = jobspec_duration (h, jobspec);
     conf->override = 0;
+    conf->reattach_finish = 0;
     conf->wait_status = 0;
     conf->mock_exception = NULL;
     conf->enabled = false;
@@ -119,9 +128,10 @@ static int init_testconf (flux_t *h, struct testconf *conf, json_t *jobspec)
         return 0;
     conf->enabled = true;
     if (json_unpack_ex (test, &err, 0,
-                        "{s?s s?i s?i s?s}",
+                        "{s?s s?i s?i s?i s?s}",
                         "run_duration", &trun,
                         "override", &conf->override,
+                        "reattach_finish", &conf->reattach_finish,
                         "wait_status", &conf->wait_status,
                         "mock_exception", &conf->mock_exception) < 0) {
         flux_log (h, LOG_ERR, "init_testconf: %s", err.text);
@@ -160,10 +170,9 @@ static void timer_cb (flux_reactor_t *r,
  *   is sent when the timer fires (simulating the exit of the final
  *   job shell.)
  */
-static int start_timer (flux_t *h, struct testexec *te)
+static int start_timer (flux_t *h, struct testexec *te, double t)
 {
     flux_reactor_t *r = flux_get_reactor (h);
-    double t = te->conf.run_duration;
 
     /*  For now, if a job duration wasn't found, complete job almost
      *   immediately.
@@ -171,7 +180,6 @@ static int start_timer (flux_t *h, struct testexec *te)
     if (t < 0.)
         t = 1.e-5;
     if (t >= 0.) {
-        char timebuf[256];
         if (t > 0.) {
             te->timer = flux_timer_watcher_create (r, t, 0., timer_cb, te);
             if (!te->timer) {
@@ -179,10 +187,11 @@ static int start_timer (flux_t *h, struct testexec *te)
                 return -1;
             }
             flux_watcher_start (te->timer);
-            snprintf (timebuf, sizeof (timebuf), "%.6fs", t);
-        } else
-            strncpy (timebuf, "inf", sizeof (timebuf));
-        jobinfo_started (te->job, "{ s:s }", "timer", timebuf);
+        }
+        if (te->job->reattach)
+            jobinfo_reattached (te->job);
+        else
+            jobinfo_started (te->job);
     }
     else
         return -1;
@@ -219,17 +228,104 @@ static int testexec_init (struct jobinfo *job)
     return 1;
 }
 
+static int testexec_reattach_starttime (struct jobinfo *job,
+                                        const char *eventlog,
+                                        time_t *startp)
+{
+    json_t *o = NULL;
+    size_t index;
+    json_t *value;
+    int rv = -1;
+
+    if (!(o = eventlog_decode (eventlog))) {
+        jobinfo_fatal_error (job, errno, "eventlog_decode");
+        goto cleanup;
+    }
+
+    json_array_foreach (o, index, value) {
+        double timestamp;
+        const char *name;
+
+        if (eventlog_entry_parse (value, &timestamp, &name, NULL) < 0) {
+            jobinfo_fatal_error (job, errno, "eventlog_entry_parse");
+            goto cleanup;
+        }
+        if (!strcmp (name, "start")) {
+            (*startp) = (time_t)timestamp;
+            break;
+        }
+    }
+
+    rv = 0;
+cleanup:
+    json_decref (o);
+    return rv;
+}
+
+static int testexec_reattach (struct testexec *te)
+{
+    const char *value;
+    flux_future_t *f = NULL;
+    time_t start = 0;
+    struct timespec now;
+    time_t runtimeleft = -1;
+    int rv = -1;
+    char ekey[256];
+
+    if (flux_job_kvs_key (ekey, sizeof (ekey), te->job->id, "eventlog") < 0) {
+        jobinfo_fatal_error (te->job, errno, "flux_job_kvs_key");
+        goto cleanup;
+    }
+    if (!(f = flux_kvs_lookup (te->job->h,
+                               NULL,
+                               0,
+                               ekey))) {
+        jobinfo_fatal_error (te->job, errno, "flux_kvs_lookup");
+        goto cleanup;
+    }
+    if (flux_kvs_lookup_get (f, &value) < 0) {
+        jobinfo_fatal_error (te->job, errno, "flux_kvs_lookup_get starttimes");
+        goto cleanup;
+    }
+    if (testexec_reattach_starttime (te->job, value, &start) < 0)
+        goto cleanup;
+    if (!te->conf.reattach_finish) {
+        /* just use seconds, we approximate runtime left */
+        clock_gettime (CLOCK_REALTIME, &now);
+        if ((now.tv_sec - start) <= te->conf.run_duration)
+            runtimeleft = (start + te->conf.run_duration) - now.tv_sec;
+    }
+    if (start_timer (te->job->h,
+                     te,
+                     runtimeleft) < 0) {
+        jobinfo_fatal_error (te->job, errno, "unable to restart timer");
+        goto cleanup;
+    }
+    rv = 0;
+cleanup:
+    flux_future_destroy (f);
+    return rv;
+}
+
 static int testexec_start (struct jobinfo *job)
 {
     struct testexec *te = job->data;
 
-    if (!te->conf.override && start_timer (job->h, te) < 0) {
-        jobinfo_fatal_error (job, errno, "unable to start test exec timer");
-        return -1;
+    if (job->reattach) {
+        if (testexec_reattach (te) < 0)
+            return -1;
     }
-    if (testconf_mock_exception (&te->conf, "run")) {
-        jobinfo_fatal_error (job, 0, "mock run exception generated");
-        return -1;
+    else {
+        if (!te->conf.override && start_timer (job->h,
+                                               te,
+                                               te->conf.run_duration) < 0) {
+            jobinfo_fatal_error (job, errno, "unable to start test exec timer");
+            return -1;
+        }
+        if (testconf_mock_exception (&te->conf, "run")) {
+            jobinfo_fatal_error (job, 0, "mock run exception generated");
+            return -1;
+        }
     }
     return 0;
 }
@@ -297,7 +393,7 @@ static void testexec_request_cb (flux_t *h,
             errno = EINVAL;
             goto error;
         }
-        if (start_timer (h, te) < 0)
+        if (start_timer (h, te, te->conf.run_duration) < 0)
             goto error;
     }
     else if (strcmp (event, "finish") == 0) {
diff --git a/src/modules/job-manager/job.h b/src/modules/job-manager/job.h
index d3d4a14332ee..a3d86ee1abfb 100644
--- a/src/modules/job-manager/job.h
+++ b/src/modules/job-manager/job.h
@@ -40,6 +40,7 @@ struct job {
     uint8_t free_pending:1; // free request sent to sched
     uint8_t has_resources:1;
     uint8_t start_pending:1;// start request sent to job-exec
+    uint8_t reattach:1;
 
     uint8_t perilog_active; // if nonzero, prolog/epilog active
 
diff --git a/src/modules/job-manager/restart.c b/src/modules/job-manager/restart.c
index b932a6b3d7d7..ec613189d877 100644
--- a/src/modules/job-manager/restart.c
+++ b/src/modules/job-manager/restart.c
@@ -266,8 +266,20 @@ int restart_from_kvs (struct job_manager *ctx)
                                 __FUNCTION__, (uintmax_t)job->id);
             }
         }
-        else if ((job->state & FLUX_JOB_STATE_RUNNING) != 0)
+        else if ((job->state & FLUX_JOB_STATE_RUNNING) != 0) {
             ctx->running_jobs++;
+            job->reattach = 1;
+            if ((job->flags & FLUX_JOB_DEBUG)) {
+                if (event_job_post_pack (ctx->event,
+                                         job,
+                                         "debug.exec-reattach-start",
+                                         0,
+                                         "{s:I}",
+                                         "id", (uintmax_t)job->id) < 0)
+                    flux_log_error (ctx->h, "%s: event_job_post_pack id=%ju",
+                                    __FUNCTION__, (uintmax_t)job->id);
+            }
+        }
         job = zhashx_next (ctx->active_jobs);
     }
     flux_log (ctx->h, LOG_INFO, "restart: %d running jobs", ctx->running_jobs);
diff --git a/src/modules/job-manager/start.c b/src/modules/job-manager/start.c
index bc7a11c8b35b..cd0084a43d33 100644
--- a/src/modules/job-manager/start.c
+++ b/src/modules/job-manager/start.c
@@ -200,8 +200,25 @@ static void start_response_cb (flux_t *h, flux_msg_handler_t *mh,
         goto error;
     }
     if (!strcmp (type, "start")) {
-        if (event_job_post_pack (ctx->event, job, "start", 0, NULL) < 0)
-            goto error_post;
+        if (job->reattach)
+            flux_log (h,
+                      LOG_ERR,
+                      "start response: id=%ju should not get start event",
+                      (uintmax_t)id);
+        else {
+            if (event_job_post_pack (ctx->event, job, "start", 0, NULL) < 0)
+                goto error_post;
+        }
+    }
+    else if (!strcmp (type, "reattached")) {
+        if ((job->flags & FLUX_JOB_DEBUG)) {
+            if (event_job_post_pack (ctx->event,
+                                     job,
+                                     "debug.exec-reattach-finish",
+                                     0,
+                                     NULL) < 0)
+                goto error_post;
+        }
     }
     else if (!strcmp (type, "release")) {
         const char *idset;
@@ -275,10 +292,11 @@ int start_send_request (struct start *start, struct job *job)
     if (!job->start_pending && start->topic != NULL) {
         if (!(msg = flux_request_encode (start->topic, NULL)))
             return -1;
-        if (flux_msg_pack (msg, "{s:I s:i s:O}",
+        if (flux_msg_pack (msg, "{s:I s:i s:O s:b}",
                                 "id", job->id,
                                 "userid", job->userid,
-                                "jobspec", job->jobspec_redacted) < 0)
+                                "jobspec", job->jobspec_redacted,
+                                "reattach", job->reattach) < 0)
             goto error;
         if (flux_send (ctx->h, msg, 0) < 0)
             goto error;
diff --git a/t/Makefile.am b/t/Makefile.am
index b317cd8cf56e..225704516186 100644
--- a/t/Makefile.am
+++ b/t/Makefile.am
@@ -47,7 +47,8 @@ clean-local:
 LONGTESTSCRIPTS = \
 	t5000-valgrind.t \
 	t3100-flux-in-flux.t \
-	t3200-instance-restart.t
+	t3200-instance-restart.t \
+	t3202-instance-restart-testexec.t
 
 # This list is included in both TESTS and dist_check_SCRIPTS.
 TESTSCRIPTS = \
diff --git a/t/t3202-instance-restart-testexec.t b/t/t3202-instance-restart-testexec.t
new file mode 100755
index 000000000000..f5e1a8dadf25
--- /dev/null
+++ b/t/t3202-instance-restart-testexec.t
@@ -0,0 +1,57 @@
+#!/bin/sh
+
+test_description='Test instance restart and still running jobs with testexec'
+
+# Append --logfile option if FLUX_TESTS_LOGFILE is set in environment:
+test -n "$FLUX_TESTS_LOGFILE" && set -- "$@" --logfile
+. `dirname $0`/sharness.sh
+
+export FLUX_INSTANCE_RESTART=t
+
+test_expect_success 'run a testexec job in persistent instance (long run)' '
+	flux start -o,--setattr=content.backing-path=$(pwd)/content.sqlite \
+	     flux mini submit \
+	       --flags=debug \
+	       --setattr=system.exec.test.run_duration=100s \
+	       hostname >id1.out
+'
+
+test_expect_success 'restart instance, reattach to running job, cancel it (long run)' '
+	flux start -o,--setattr=content.backing-path=$(pwd)/content.sqlite \
+	     sh -c "flux job eventlog $(cat id1.out) > eventlog_long1.out; \
+		    flux jobs -n > jobs_long1.out; \
+		    flux job cancel $(cat id1.out)" &&
+	grep "reattach-start" eventlog_long1.out &&
+	grep "reattach-finish" eventlog_long1.out &&
+	grep $(cat id1.out) jobs_long1.out
+'
+
+test_expect_success 'restart instance, job completed (long run)' '
+	flux start -o,--setattr=content.backing-path=$(pwd)/content.sqlite \
+	     sh -c "flux job eventlog $(cat id1.out) > eventlog_long2.out; \
+		    flux jobs -n > jobs_long2.out" &&
+	grep "finish" eventlog_long2.out | grep status &&
+	test_must_fail grep $(cat id1.out) jobs_long2.out
+'
+
+# reattach_finish will indicate to testexcec that the job finished
+# right after reattach, emulating a job that finished before the
+# instance restarted
+test_expect_success 'run a testexec job in persistent instance (exit run)' '
+	flux start -o,--setattr=content.backing-path=$(pwd)/content.sqlite \
+	     flux mini submit \
+	       --flags=debug \
+	       --setattr=system.exec.test.reattach_finish=1 \
+	       --setattr=system.exec.test.run_duration=100s \
+	     hostname >id2.out
+'
+
+test_expect_success 'restart instance, reattach to running job, its finished (exit run)' '
+	flux start -o,--setattr=content.backing-path=$(pwd)/content.sqlite \
+	     sh -c "flux job eventlog $(cat id2.out) > eventlog_exit1.out" &&
+	grep "reattach-start" eventlog_exit1.out &&
+	grep "reattach-finish" eventlog_exit1.out &&
+	grep "finish" eventlog_exit1.out | grep status
+'
+
+test_done