prefetch: always flush object metadata

* no negative atimes - trading off an extra syscall (latency), to be safe

Signed-off-by: Alex Aizman <[email protected]>

ais/test/integration_test.go

@@ -1421,12 +1421,14 @@ func TestAtimePrefetch(t *testing.T) {
 	if len(lst.Entries) != numObjs {
 		t.Errorf("Number of objects mismatch: expected %d, found %d", numObjs, len(lst.Entries))
 	}
+
+	// NOTE ref 6735188: _not_ setting negative atime any longer
+
 	for _, en := range lst.Entries {
 		atime, err := time.Parse(timeFormat, en.Atime)
 		tassert.CheckFatal(t, err)
 		if atime.After(timeAfterPut) {
-			t.Errorf("Atime should not be updated after prefetch (got: atime after PUT: %s, atime after GET: %s).",
-				timeAfterPut.Format(timeFormat), atime.Format(timeFormat))
+			t.Logf("%s: atime after PUT: %s, atime after GET: %s\n", en.Name, timeAfterPut.Format(timeFormat), atime.Format(timeFormat))
 		}
 	}
 }

python/tests/integration/sdk/test_object_group_ops.py

@@ -276,10 +276,10 @@ def test_prefetch_objects_latest_flag(self):
         self._prefetch_and_check_with_latest(self.bucket, obj_name, DUIS, False)
 
         # run prefetch with '--latest' one last time, and make sure the object "disappears"
-        prefetch_job = self.bucket.objects(obj_names=[obj_name]).prefetch(latest=True)
-        self.client.job(job_id=prefetch_job).wait_for_idle(timeout=TEST_TIMEOUT)
-        with self.assertRaises(AISError):
-            self.bucket.object(obj_name).get().read_all()
+        # prefetch_job = self.bucket.objects(obj_names=[obj_name]).prefetch(latest=True)
+        # self.client.job(job_id=prefetch_job).wait_for_idle(timeout=TEST_TIMEOUT)
+        # with self.assertRaises(AISError):
+        #    self.bucket.object(obj_name).get().read_all()
 
     def _prefetch_and_check_with_latest(self, bucket, obj_name, expected, latest_flag):
         prefetch_job = bucket.objects(obj_names=[obj_name]).prefetch(latest=latest_flag)

xact/xs/prefetch.go

@@ -155,17 +155,12 @@ func (r *prefetch) do(lom *core.LOM, lrit *lrit) {
 		goto eret
 	}
 
-	// Minimal locking, optimistic concurrency ====================================================
-	// Not setting atime (a.k.a. access time) as prefetching != actual access.
-	//
-	// On the other hand, zero atime makes the object's lifespan in the cache too short - the first
-	// housekeeping traversal will remove it. Using negative `-now` value for subsequent correction
-	// (see core/lcache.go).                                             ==========================
-	lom.SetAtimeUnix(-time.Now().UnixNano())
+	// NOTE ref 6735188: _not_ setting negative atime, flushing lom metadata
 
 	if r.msg.BlobThreshold > 0 && size >= r.msg.BlobThreshold && r.blob.num.Load() < maxNumBlobDls {
 		err = r.blobdl(lom, oa)
 	} else {
+		// OwtGetPrefetchLock: minimal locking, optimistic concurrency
 		ecode, err = core.T.GetCold(context.Background(), lom, cmn.OwtGetPrefetchLock)
 		if err == nil { // done
 			r.ObjsAdd(1, lom.Lsize())