Golang keys.RangeMetaKey函數代碼示例

// TestReplicateRange verifies basic replication functionality by creating two stores
// and a range, replicating the range to the second store, and reading its data there.
func TestReplicateRange(t *testing.T) {
	defer leaktest.AfterTest(t)
	mtc := multiTestContext{}
	mtc.Start(t, 2)
	defer mtc.Stop()

	// Issue a command on the first node before replicating.
	incArgs, incResp := incrementArgs([]byte("a"), 5, 1, mtc.stores[0].StoreID())
	if err := mtc.stores[0].ExecuteCmd(context.Background(), proto.Call{Args: incArgs, Reply: incResp}); err != nil {
		t.Fatal(err)
	}

	rng, err := mtc.stores[0].GetRange(1)
	if err != nil {
		t.Fatal(err)
	}

	if err := rng.ChangeReplicas(proto.ADD_REPLICA,
		proto.Replica{
			NodeID:  mtc.stores[1].Ident.NodeID,
			StoreID: mtc.stores[1].Ident.StoreID,
		}); err != nil {
		t.Fatal(err)
	}
	// Verify no intent remains on range descriptor key.
	key := keys.RangeDescriptorKey(rng.Desc().StartKey)
	desc := proto.RangeDescriptor{}
	if ok, err := engine.MVCCGetProto(mtc.stores[0].Engine(), key, mtc.stores[0].Clock().Now(), true, nil, &desc); !ok || err != nil {
		t.Fatalf("fetching range descriptor yielded %t, %s", ok, err)
	}
	// Verify that in time, no intents remain on meta addressing
	// keys, and that range descriptor on the meta records is correct.
	util.SucceedsWithin(t, 1*time.Second, func() error {
		meta2 := keys.RangeMetaKey(proto.KeyMax)
		meta1 := keys.RangeMetaKey(meta2)
		for _, key := range []proto.Key{meta2, meta1} {
			metaDesc := proto.RangeDescriptor{}
			if ok, err := engine.MVCCGetProto(mtc.stores[0].Engine(), key, mtc.stores[0].Clock().Now(), true, nil, &metaDesc); !ok || err != nil {
				return util.Errorf("failed to resolve %s", key)
			}
			if !reflect.DeepEqual(metaDesc, desc) {
				return util.Errorf("descs not equal: %+v != %+v", metaDesc, desc)
			}
		}
		return nil
	})

	// Verify that the same data is available on the replica.
	util.SucceedsWithin(t, 1*time.Second, func() error {
		getArgs, getResp := getArgs([]byte("a"), 1, mtc.stores[1].StoreID())
		getArgs.ReadConsistency = proto.INCONSISTENT
		if err := mtc.stores[1].ExecuteCmd(context.Background(), proto.Call{Args: getArgs, Reply: getResp}); err != nil {
			return util.Errorf("failed to read data")
		}
		if v := mustGetInteger(getResp.Value); v != 5 {
			return util.Errorf("failed to read correct data: %d", v)
		}
		return nil
	})
}

func TestRangeCacheAssumptions(t *testing.T) {
	defer leaktest.AfterTest(t)
	expKeyMin := keys.RangeMetaKey(keys.RangeMetaKey(keys.RangeMetaKey(proto.Key("test"))))
	if !bytes.Equal(expKeyMin, proto.KeyMin) {
		t.Fatalf("RangeCache relies on RangeMetaKey returning KeyMin after two levels, but got %s", expKeyMin)
	}
}

// TestRangeCacheClearOverlappingMeta prevents regression of a bug which caused
// a panic when clearing overlapping descriptors for [KeyMin, Meta2Key). The
// issue was that when attempting to clear out descriptors which were subsumed
// by the above range, an iteration over the corresponding meta keys was
// performed, with the left endpoint excluded. This exclusion was incorrect: it
// first incremented the start key (KeyMin) and then formed the meta key; for
// KeyMin this leads to Meta2Prefix\x00. For the above EndKey, the meta key is
// a Meta1key which sorts before Meta2Prefix\x00, causing a panic. The fix was
// simply to increment the meta key for StartKey, not StartKey itself.
func TestRangeCacheClearOverlappingMeta(t *testing.T) {
	defer leaktest.AfterTest(t)

	firstDesc := &proto.RangeDescriptor{
		StartKey: proto.KeyMin,
		EndKey:   proto.Key("zzz"),
	}
	restDesc := &proto.RangeDescriptor{
		StartKey: firstDesc.StartKey,
		EndKey:   proto.KeyMax,
	}

	cache := newRangeDescriptorCache(nil, 2<<10)
	cache.rangeCache.Add(rangeCacheKey(keys.RangeMetaKey(firstDesc.EndKey)),
		firstDesc)
	cache.rangeCache.Add(rangeCacheKey(keys.RangeMetaKey(restDesc.EndKey)),
		restDesc)

	// Add new range, corresponding to splitting the first range at a meta key.
	metaSplitDesc := &proto.RangeDescriptor{
		StartKey: proto.KeyMin,
		EndKey:   proto.Key(keys.RangeMetaKey(proto.Key("foo"))),
	}
	func() {
		defer func() {
			if r := recover(); r != nil {
				t.Fatalf("invocation of clearOverlappingCachedRangeDescriptors panicked: %v", r)
			}
		}()
		cache.clearOverlappingCachedRangeDescriptors(metaSplitDesc.EndKey, keys.RangeMetaKey(metaSplitDesc.EndKey), metaSplitDesc)
	}()
}

// clearOverlappingCachedRangeDescriptors looks up and clears any
// cache entries which overlap the specified key or descriptor.
func (rdc *rangeDescriptorCache) clearOverlappingCachedRangeDescriptors(key, metaKey proto.Key, desc *proto.RangeDescriptor) {
	if desc.StartKey.Equal(desc.EndKey) { // True for some unittests.
		return
	}
	// Clear out any descriptors which subsume the key which we're going
	// to cache. For example, if an existing KeyMin->KeyMax descriptor
	// should be cleared out in favor of a KeyMin->"m" descriptor.
	k, v, ok := rdc.rangeCache.Ceil(rangeCacheKey(metaKey))
	if ok {
		descriptor := v.(*proto.RangeDescriptor)
		addrKey := keys.KeyAddress(key)
		if !addrKey.Less(descriptor.StartKey) && !descriptor.EndKey.Less(addrKey) {
			if log.V(1) {
				log.Infof("clearing overlapping descriptor: key=%s desc=%s", k, descriptor)
			}
			rdc.rangeCache.Del(k.(rangeCacheKey))
		}
	}
	// Also clear any descriptors which are subsumed by the one we're
	// going to cache. This could happen on a merge (and also happens
	// when there's a lot of concurrency). Iterate from StartKey.Next().
	rdc.rangeCache.DoRange(func(k, v interface{}) {
		if log.V(1) {
			log.Infof("clearing subsumed descriptor: key=%s desc=%s", k, v.(*proto.RangeDescriptor))
		}
		rdc.rangeCache.Del(k.(rangeCacheKey))
	}, rangeCacheKey(keys.RangeMetaKey(desc.StartKey.Next())),
		rangeCacheKey(keys.RangeMetaKey(desc.EndKey)))
}

// getCachedRangeDescriptorLocked is a helper function to retrieve the
// descriptor of the range which contains the given key, if present in the
// cache. It is assumed that the caller holds a read lock on rdc.rangeCacheMu.
func (rdc *rangeDescriptorCache) getCachedRangeDescriptorLocked(key proto.Key, inclusive bool) (
	rangeCacheKey, *proto.RangeDescriptor) {
	// The cache is indexed using the end-key of the range, but the
	// end-key is non-inclusive by default.
	var metaKey proto.Key
	if !inclusive {
		metaKey = keys.RangeMetaKey(key.Next())
	} else {
		metaKey = keys.RangeMetaKey(key)
	}

	k, v, ok := rdc.rangeCache.Ceil(rangeCacheKey(metaKey))
	if !ok {
		return nil, nil
	}
	metaEndKey := k.(rangeCacheKey)
	rd := v.(*proto.RangeDescriptor)

	// Check that key actually belongs to the range.
	if !rd.ContainsKey(key) {
		// The key is the EndKey and we're inclusive, so just return the range descriptor.
		if inclusive && key.Equal(rd.EndKey) {
			return metaEndKey, rd
		}
		return nil, nil
	}

	// The key is the StartKey, but we're inclusive and thus need to return the
	// previous range descriptor, but it is not in the cache yet.
	if inclusive && key.Equal(rd.StartKey) {
		return nil, nil
	}
	return metaEndKey, rd
}

// TestRangeSplitMeta executes various splits (including at meta addressing)
// and checks that all created intents are resolved. This includes both intents
// which are resolved synchronously with EndTransaction and via RPC.
func TestRangeSplitMeta(t *testing.T) {
	defer leaktest.AfterTest(t)
	s := createTestDB(t)
	defer s.Stop()

	splitKeys := []roachpb.Key{roachpb.Key("G"), keys.RangeMetaKey(roachpb.Key("F")),
		keys.RangeMetaKey(roachpb.Key("K")), keys.RangeMetaKey(roachpb.Key("H"))}

	// Execute the consecutive splits.
	for _, splitKey := range splitKeys {
		log.Infof("starting split at key %q...", splitKey)
		if err := s.DB.AdminSplit(splitKey); err != nil {
			t.Fatal(err)
		}
		log.Infof("split at key %q complete", splitKey)
	}

	if err := util.IsTrueWithin(func() bool {
		if _, _, err := engine.MVCCScan(s.Eng, keys.LocalMax, roachpb.KeyMax, 0, roachpb.MaxTimestamp, true, nil); err != nil {
			log.Infof("mvcc scan should be clean: %s", err)
			return false
		}
		return true
	}, 500*time.Millisecond); err != nil {
		t.Error("failed to verify no dangling intents within 500ms")
	}
}

// TestReplicateRange verifies basic replication functionality by creating two stores
// and a range, replicating the range to the second store, and reading its data there.
func TestReplicateRange(t *testing.T) {
	defer leaktest.AfterTest(t)
	mtc := startMultiTestContext(t, 2)
	defer mtc.Stop()

	// Issue a command on the first node before replicating.
	incArgs := incrementArgs([]byte("a"), 5)
	if _, err := client.SendWrapped(rg1(mtc.stores[0]), nil, &incArgs); err != nil {
		t.Fatal(err)
	}

	rng, err := mtc.stores[0].GetReplica(1)
	if err != nil {
		t.Fatal(err)
	}

	if err := rng.ChangeReplicas(roachpb.ADD_REPLICA,
		roachpb.ReplicaDescriptor{
			NodeID:  mtc.stores[1].Ident.NodeID,
			StoreID: mtc.stores[1].Ident.StoreID,
		}, rng.Desc()); err != nil {
		t.Fatal(err)
	}
	// Verify no intent remains on range descriptor key.
	key := keys.RangeDescriptorKey(rng.Desc().StartKey)
	desc := roachpb.RangeDescriptor{}
	if ok, err := engine.MVCCGetProto(mtc.stores[0].Engine(), key, mtc.stores[0].Clock().Now(), true, nil, &desc); !ok || err != nil {
		t.Fatalf("fetching range descriptor yielded %t, %s", ok, err)
	}
	// Verify that in time, no intents remain on meta addressing
	// keys, and that range descriptor on the meta records is correct.
	util.SucceedsWithin(t, 1*time.Second, func() error {
		meta2 := keys.Addr(keys.RangeMetaKey(roachpb.RKeyMax))
		meta1 := keys.Addr(keys.RangeMetaKey(meta2))
		for _, key := range []roachpb.RKey{meta2, meta1} {
			metaDesc := roachpb.RangeDescriptor{}
			if ok, err := engine.MVCCGetProto(mtc.stores[0].Engine(), key.AsRawKey(), mtc.stores[0].Clock().Now(), true, nil, &metaDesc); !ok || err != nil {
				return util.Errorf("failed to resolve %s", key.AsRawKey())
			}
			if !reflect.DeepEqual(metaDesc, desc) {
				return util.Errorf("descs not equal: %+v != %+v", metaDesc, desc)
			}
		}
		return nil
	})

	// Verify that the same data is available on the replica.
	util.SucceedsWithin(t, replicaReadTimeout, func() error {
		getArgs := getArgs([]byte("a"))
		if reply, err := client.SendWrappedWith(rg1(mtc.stores[1]), nil, roachpb.Header{
			ReadConsistency: roachpb.INCONSISTENT,
		}, &getArgs); err != nil {
			return util.Errorf("failed to read data: %s", err)
		} else if e, v := int64(5), mustGetInt(reply.(*roachpb.GetResponse).Value); v != e {
			return util.Errorf("failed to read correct data: expected %d, got %d", e, v)
		}
		return nil
	})
}

// EvictCachedRangeDescriptor will evict any cached range descriptors
// for the given key. It is intended that this method be called from a
// consumer of rangeDescriptorCache if the returned range descriptor is
// discovered to be stale.
// seenDesc should always be passed in and is used as the basis of a
// compare-and-evict (as pointers); if it is nil, eviction is unconditional
// but a warning will be logged.
func (rdc *rangeDescriptorCache) EvictCachedRangeDescriptor(descKey proto.Key, seenDesc *proto.RangeDescriptor) {
	if seenDesc == nil {
		log.Warningf("compare-and-evict for key %s with nil descriptor; clearing unconditionally", descKey)
	}

	rdc.rangeCacheMu.Lock()
	defer rdc.rangeCacheMu.Unlock()

	rngKey, cachedDesc := rdc.getCachedRangeDescriptorLocked(descKey)
	// Note that we're doing a "compare-and-erase": If seenDesc is not nil,
	// we want to clean the cache only if it equals the cached range
	// descriptor as a pointer. If not, then likely some other caller
	// already evicted previously, and we can save work by not doing it
	// again (which would prompt another expensive lookup).
	if seenDesc != nil && seenDesc != cachedDesc {
		return
	}

	for !bytes.Equal(descKey, proto.KeyMin) {
		if log.V(2) {
			log.Infof("evict cached descriptor: key=%s desc=%s\n%s", descKey, cachedDesc, rdc.stringLocked())
		} else if log.V(1) {
			log.Infof("evict cached descriptor: key=%s desc=%s", descKey, cachedDesc)
		}
		rdc.rangeCache.Del(rngKey)

		// Retrieve the metadata range key for the next level of metadata, and
		// evict that key as well. This loop ends after the meta1 range, which
		// returns KeyMin as its metadata key.
		descKey = keys.RangeMetaKey(descKey)
		rngKey, cachedDesc = rdc.getCachedRangeDescriptorLocked(descKey)
	}
}

func metaKey(key roachpb.RKey) []byte {
	rk, err := keys.Addr(keys.RangeMetaKey(key))
	if err != nil {
		panic(err)
	}
	return rk
}

func runLsRanges(cmd *cobra.Command, args []string) {
	if len(args) > 1 {
		mustUsage(cmd)
		return
	}
	var startKey proto.Key
	if len(args) >= 1 {
		startKey = keys.RangeMetaKey(proto.Key(args[0]))
	} else {
		startKey = keys.Meta2Prefix
	}

	kvDB, stopper := makeDBClient()
	defer stopper.Stop()
	rows, err := kvDB.Scan(startKey, keys.Meta2Prefix.PrefixEnd(), maxResults)
	if err != nil {
		fmt.Fprintf(os.Stderr, "scan failed: %s\n", err)
		osExit(1)
		return
	}

	for _, row := range rows {
		desc := &proto.RangeDescriptor{}
		if err := row.ValueProto(desc); err != nil {
			fmt.Fprintf(os.Stderr, "%s: unable to unmarshal range descriptor\n", row.Key)
			continue
		}
		fmt.Printf("%s-%s [%d]\n", desc.StartKey, desc.EndKey, desc.RangeID)
		for i, replica := range desc.Replicas {
			fmt.Printf("\t%d: node-id=%d store-id=%d\n",
				i, replica.NodeID, replica.StoreID)
		}
	}
	fmt.Printf("%d result(s)\n", len(rows))
}

// rangeAddressing updates or deletes the range addressing metadata
// for the range specified by desc. The action to take is specified by
// the supplied metaAction function.
//
// The rules for meta1 and meta2 records are as follows:
//
//  1. If desc.StartKey or desc.EndKey is meta1:
//     - ERROR
//  2. If desc.EndKey is meta2:
//     - meta1(desc.EndKey)
//  3. If desc.EndKey is normal user key:
//     - meta2(desc.EndKey)
//     3a. If desc.StartKey is KeyMin or meta2:
//         - meta1(KeyMax)
func rangeAddressing(b *client.Batch, desc *roachpb.RangeDescriptor, action metaAction) error {
	// 1. handle illegal case of start or end key being meta1.
	if bytes.HasPrefix(desc.EndKey, keys.Meta1Prefix) ||
		bytes.HasPrefix(desc.StartKey, keys.Meta1Prefix) {
		return errors.Errorf("meta1 addressing records cannot be split: %+v", desc)
	}

	// Note that both cases 2 and 3 are handled by keys.RangeMetaKey.
	//
	// 2. the case of the range ending with a meta2 prefix. This means
	// the range is full of meta2. We must update the relevant meta1
	// entry pointing to the end of this range.
	//
	// 3. the range ends with a normal user key, so we must update the
	// relevant meta2 entry pointing to the end of this range.
	action(b, keys.RangeMetaKey(desc.EndKey), desc)

	if !bytes.HasPrefix(desc.EndKey, keys.Meta2Prefix) {
		// 3a. the range starts with KeyMin or a meta2 addressing record,
		// update the meta1 entry for KeyMax.
		if bytes.Equal(desc.StartKey, roachpb.RKeyMin) ||
			bytes.HasPrefix(desc.StartKey, keys.Meta2Prefix) {
			action(b, keys.Meta1KeyMax, desc)
		}
	}
	return nil
}

func (db *testDescriptorDB) getRangeDescriptors(key proto.Key,
	options lookupOptions) ([]proto.RangeDescriptor, error) {
	db.lookupCount++
	metadataKey := keys.RangeMetaKey(key)

	var err error

	// Recursively call into cache as the real DB would, terminating recursion
	// when a meta1key is encountered.
	if len(metadataKey) > 0 && !bytes.HasPrefix(metadataKey, keys.Meta1Prefix) {
		_, err = db.cache.LookupRangeDescriptor(metadataKey, options)
	}
	return db.getDescriptor(key), err
}

// getCachedRangeDescriptorLocked is a helper function to retrieve the
// descriptor of the range which contains the given key, if present in the
// cache. It is assumed that the caller holds a read lock on rdc.rangeCacheMu.
func (rdc *rangeDescriptorCache) getCachedRangeDescriptorLocked(key proto.Key, isReverse bool) (
	rangeCacheKey, *proto.RangeDescriptor) {
	// The cache is indexed using the end-key of the range, but the
	// end-key is non-inclusive.
	var metaKey proto.Key
	if !isReverse {
		// If it is not reverse scan, we access the cache using key.Next().
		metaKey = keys.RangeMetaKey(key.Next())
	} else {
		// Because reverse scan request is begining at end key(exclusive),so we
		// access the cache using key directly.
		metaKey = keys.RangeMetaKey(key)
	}

	k, v, ok := rdc.rangeCache.Ceil(rangeCacheKey(metaKey))
	if !ok {
		return nil, nil
	}
	metaEndKey := k.(rangeCacheKey)
	rd := v.(*proto.RangeDescriptor)

	// Check that key actually belongs to the range.
	if !rd.ContainsKey(keys.KeyAddress(key)) {
		// The key is the EndKey of the range in reverse scan, just return the range descriptor.
		if isReverse && key.Equal(rd.EndKey) {
			return metaEndKey, rd
		}
		return nil, nil
	}

	// The key is the StartKey of the range in reverse scan. We need to return the previous range
	// descriptor, but it is not in the cache yet.
	if isReverse && key.Equal(rd.StartKey) {
		return nil, nil
	}
	return metaEndKey, rd
}

// LookupRange returns the descriptor of the range containing key.
func (tc *TestCluster) LookupRange(key roachpb.Key) (roachpb.RangeDescriptor, error) {
	rangeLookupReq := roachpb.RangeLookupRequest{
		Span: roachpb.Span{
			Key: keys.RangeMetaKey(keys.MustAddr(key)),
		},
		MaxRanges:       1,
		ConsiderIntents: false,
	}
	resp, pErr := client.SendWrapped(tc.Servers[0].GetDistSender(), nil, &rangeLookupReq)
	if pErr != nil {
		return roachpb.RangeDescriptor{}, errors.Errorf(
			"%q: lookup range unexpected error: %s", key, pErr)
	}
	return resp.(*roachpb.RangeLookupResponse).Ranges[0], nil
}

func (c *cluster) lookupRange(nodeIdx int, key roachpb.Key) (*roachpb.RangeDescriptor, error) {
	req := &roachpb.RangeLookupRequest{
		Span: roachpb.Span{
			Key: keys.RangeMetaKey(keys.MustAddr(key)),
		},
		MaxRanges:       1,
		ConsiderIntents: false,
	}
	sender := c.clients[nodeIdx].GetSender()
	resp, pErr := client.SendWrapped(sender, nil, req)
	if pErr != nil {
		return nil, errors.Errorf("%s: lookup range: %s", key, pErr)
	}
	return &resp.(*roachpb.RangeLookupResponse).Ranges[0], nil
}