// prev gives the right boundary of the union of all requests which don't
// affect keys larger than the given key.
// TODO(tschottdorf): again, better on BatchRequest itself, but can't pull
// 'keys' into 'roachpb'.
func prev(ba roachpb.BatchRequest, k roachpb.RKey) roachpb.RKey {
candidate := roachpb.RKeyMin
for _, union := range ba.Requests {
h := union.GetInner().Header()
addr := keys.Addr(h.Key)
eAddr := keys.Addr(h.EndKey)
if len(eAddr) == 0 {
// Can probably avoid having to compute Next() here if
// we're in the mood for some more complexity.
eAddr = addr.Next()
}
if !eAddr.Less(k) {
if !k.Less(addr) {
// Range contains k, so won't be able to go lower.
return k
}
// Range is disjoint from [KeyMin,k).
continue
}
// We want the largest surviving candidate.
if candidate.Less(addr) {
candidate = addr
}
}
return candidate
}
// next gives the left boundary of the union of all requests which don't
// affect keys less than the given key.
// TODO(tschottdorf): again, better on BatchRequest itself, but can't pull
// 'keys' into 'proto'.
func next(ba roachpb.BatchRequest, k roachpb.RKey) (roachpb.RKey, error) {
candidate := roachpb.RKeyMax
for _, union := range ba.Requests {
h := union.GetInner().Header()
addr, err := keys.Addr(h.Key)
if err != nil {
return nil, err
}
if addr.Less(k) {
eAddr, err := keys.AddrUpperBound(h.EndKey)
if err != nil {
return nil, err
}
if k.Less(eAddr) {
// Starts below k, but continues beyond. Need to stay at k.
return k, nil
}
// Affects only [KeyMin,k).
continue
}
// We want the smallest of the surviving candidates.
if addr.Less(candidate) {
candidate = addr
}
}
return candidate, nil
}
// RangeLookup dispatches an RangeLookup request for the given
// metadata key to the replicas of the given range. Note that we allow
// inconsistent reads when doing range lookups for efficiency. Getting
// stale data is not a correctness problem but instead may
// infrequently result in additional latency as additional range
// lookups may be required. Note also that rangeLookup bypasses the
// DistSender's Send() method, so there is no error inspection and
// retry logic here; this is not an issue since the lookup performs a
// single inconsistent read only.
func (ds *DistSender) RangeLookup(key roachpb.RKey, desc *roachpb.RangeDescriptor, considerIntents, useReverseScan bool) ([]roachpb.RangeDescriptor, *roachpb.Error) {
ba := roachpb.BatchRequest{}
ba.ReadConsistency = roachpb.INCONSISTENT
ba.Add(&roachpb.RangeLookupRequest{
Span: roachpb.Span{
// We can interpret the RKey as a Key here since it's a metadata
// lookup; those are never local.
Key: key.AsRawKey(),
},
MaxRanges: ds.rangeLookupMaxRanges,
ConsiderIntents: considerIntents,
Reverse: useReverseScan,
})
replicas := newReplicaSlice(ds.gossip, desc)
// TODO(tschottdorf) consider a Trace here, potentially that of the request
// that had the cache miss and waits for the result.
br, err := ds.sendRPC(nil /* Trace */, desc.RangeID, replicas, rpc.OrderRandom, ba)
if err != nil {
return nil, err
}
if br.Error != nil {
return nil, br.Error
}
return br.Responses[0].GetInner().(*roachpb.RangeLookupResponse).Ranges, nil
}
// RangeLookup implements the RangeDescriptorDB interface.
// RangeLookup dispatches a RangeLookup request for the given metadata
// key to the replicas of the given range. Note that we allow
// inconsistent reads when doing range lookups for efficiency. Getting
// stale data is not a correctness problem but instead may
// infrequently result in additional latency as additional range
// lookups may be required. Note also that rangeLookup bypasses the
// DistSender's Send() method, so there is no error inspection and
// retry logic here; this is not an issue since the lookup performs a
// single inconsistent read only.
func (ds *DistSender) RangeLookup(
key roachpb.RKey, desc *roachpb.RangeDescriptor, considerIntents, useReverseScan bool,
) ([]roachpb.RangeDescriptor, []roachpb.RangeDescriptor, *roachpb.Error) {
ba := roachpb.BatchRequest{}
ba.ReadConsistency = roachpb.INCONSISTENT
ba.Add(&roachpb.RangeLookupRequest{
Span: roachpb.Span{
// We can interpret the RKey as a Key here since it's a metadata
// lookup; those are never local.
Key: key.AsRawKey(),
},
MaxRanges: ds.rangeLookupMaxRanges,
ConsiderIntents: considerIntents,
Reverse: useReverseScan,
})
replicas := newReplicaSlice(ds.gossip, desc)
replicas.Shuffle()
// TODO(tschottdorf): Ideally we would use the trace of the request which
// caused this lookup.
_ = context.TODO()
br, err := ds.sendRPC(ds.Ctx, desc.RangeID, replicas, ba)
if err != nil {
return nil, nil, roachpb.NewError(err)
}
if br.Error != nil {
return nil, nil, br.Error
}
resp := br.Responses[0].GetInner().(*roachpb.RangeLookupResponse)
return resp.Ranges, resp.PrefetchedRanges, nil
}
// prev gives the right boundary of the union of all requests which don't
// affect keys larger than the given key.
// TODO(tschottdorf): again, better on BatchRequest itself, but can't pull
// 'keys' into 'roachpb'.
func prev(ba roachpb.BatchRequest, k roachpb.RKey) (roachpb.RKey, error) {
candidate := roachpb.RKeyMin
for _, union := range ba.Requests {
h := union.GetInner().Header()
addr, err := keys.Addr(h.Key)
if err != nil {
return nil, err
}
eAddr, err := keys.AddrUpperBound(h.EndKey)
if err != nil {
return nil, err
}
if len(eAddr) == 0 {
eAddr = addr.Next()
}
if !eAddr.Less(k) {
if !k.Less(addr) {
// Range contains k, so won't be able to go lower.
return k, nil
}
// Range is disjoint from [KeyMin,k).
continue
}
// We want the largest surviving candidate.
if candidate.Less(addr) {
candidate = addr
}
}
return candidate, nil
}
// 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.rangeCache.
func (rdc *rangeDescriptorCache) getCachedRangeDescriptorLocked(key roachpb.RKey, inclusive bool) (rangeCacheKey, *roachpb.RangeDescriptor, error) {
// The cache is indexed using the end-key of the range, but the
// end-key is non-inclusive by default.
var metaKey roachpb.RKey
var err error
if !inclusive {
metaKey, err = meta(key.Next())
} else {
metaKey, err = meta(key)
}
if err != nil {
return nil, nil, err
}
k, v, ok := rdc.rangeCache.cache.Ceil(rangeCacheKey(metaKey))
if !ok {
return nil, nil, nil
}
metaEndKey := k.(rangeCacheKey)
rd := v.(*roachpb.RangeDescriptor)
// Return nil if the key does not belong to the range.
if (!inclusive && !rd.ContainsKey(key)) || (inclusive && !rd.ContainsExclusiveEndKey(key)) {
return nil, nil, nil
}
return metaEndKey, rd, nil
}
// getDescriptors looks up the range descriptor to use for a query over the
// key range [from,to), with the given lookupOptions. The range descriptor
// which contains the range in which the request should start its query is
// returned first; the returned bool is true in case the given range reaches
// outside the first descriptor.
// In case either of the descriptors is discovered stale, the returned closure
// should be called; it evicts the cache appropriately.
// Note that `from` and `to` are not necessarily Key and EndKey from a
// RequestHeader; it's assumed that they've been translated to key addresses
// already (via KeyAddress).
func (ds *DistSender) getDescriptors(from, to roachpb.RKey, options lookupOptions) (*roachpb.RangeDescriptor, bool, func(), *roachpb.Error) {
var desc *roachpb.RangeDescriptor
var err error
var descKey roachpb.RKey
if !options.useReverseScan {
descKey = from
} else {
descKey = to
}
desc, err = ds.rangeCache.LookupRangeDescriptor(descKey, options)
if err != nil {
return nil, false, nil, roachpb.NewError(err)
}
// Checks whether need to get next range descriptor. If so, returns true.
needAnother := func(desc *roachpb.RangeDescriptor, isReverse bool) bool {
if isReverse {
return from.Less(desc.StartKey)
}
return desc.EndKey.Less(to)
}
evict := func() {
ds.rangeCache.EvictCachedRangeDescriptor(descKey, desc, options.useReverseScan)
}
return desc, needAnother(desc, options.useReverseScan), evict, nil
}
// RangeLookup dispatches an RangeLookup request for the given
// metadata key to the replicas of the given range. Note that we allow
// inconsistent reads when doing range lookups for efficiency. Getting
// stale data is not a correctness problem but instead may
// infrequently result in additional latency as additional range
// lookups may be required. Note also that rangeLookup bypasses the
// DistSender's Send() method, so there is no error inspection and
// retry logic here; this is not an issue since the lookup performs a
// single inconsistent read only.
func (ds *DistSender) RangeLookup(key roachpb.RKey, desc *roachpb.RangeDescriptor, considerIntents, useReverseScan bool) ([]roachpb.RangeDescriptor, *roachpb.Error) {
ba := roachpb.BatchRequest{}
ba.ReadConsistency = roachpb.INCONSISTENT
ba.Add(&roachpb.RangeLookupRequest{
Span: roachpb.Span{
// We can interpret the RKey as a Key here since it's a metadata
// lookup; those are never local.
Key: key.AsRawKey(),
},
MaxRanges: ds.rangeLookupMaxRanges,
ConsiderIntents: considerIntents,
Reverse: useReverseScan,
})
replicas := newReplicaSlice(ds.gossip, desc)
trace := ds.Tracer.StartSpan("range lookup")
defer trace.Finish()
// TODO(tschottdorf): Ideally we would use the trace of the request which
// caused this lookup instead of a new one.
br, err := ds.sendRPC(trace, desc.RangeID, replicas, orderRandom, ba)
if err != nil {
return nil, err
}
if br.Error != nil {
return nil, br.Error
}
return br.Responses[0].GetInner().(*roachpb.RangeLookupResponse).Ranges, nil
}
// lookupReplica looks up replica by key [range]. Lookups are done
// by consulting each store in turn via Store.LookupRange(key).
// Returns RangeID and replica on success; RangeKeyMismatch error
// if not found.
// This is only for testing usage; performance doesn't matter.
func (ls *Stores) lookupReplica(start, end roachpb.RKey) (rangeID roachpb.RangeID, replica *roachpb.ReplicaDescriptor, pErr *roachpb.Error) {
ls.mu.RLock()
defer ls.mu.RUnlock()
var rng *Replica
for _, store := range ls.storeMap {
rng = store.LookupReplica(start, end)
if rng == nil {
if tmpRng := store.LookupReplica(start, nil); tmpRng != nil {
log.Warningf(fmt.Sprintf("range not contained in one range: [%s,%s), but have [%s,%s)", start, end, tmpRng.Desc().StartKey, tmpRng.Desc().EndKey))
}
continue
}
if replica == nil {
rangeID = rng.RangeID
replica = rng.GetReplica()
continue
}
// Should never happen outside of tests.
return 0, nil, roachpb.NewErrorf(
"range %+v exists on additional store: %+v", rng, store)
}
if replica == nil {
pErr = roachpb.NewError(roachpb.NewRangeKeyMismatchError(start.AsRawKey(), end.AsRawKey(), nil))
}
return rangeID, replica, pErr
}
// clearOverlappingCachedRangeDescriptors looks up and clears any
// cache entries which overlap the specified key or descriptor.
func (rdc *rangeDescriptorCache) clearOverlappingCachedRangeDescriptors(key, metaKey roachpb.RKey, desc *roachpb.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.(*roachpb.RangeDescriptor)
if !key.Less(descriptor.StartKey) && !descriptor.EndKey.Less(key) {
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 the range meta key
// after RangeMetaKey(desc.StartKey) to the range meta key for desc.EndKey.
rdc.rangeCache.DoRange(func(k, v interface{}) {
if log.V(1) {
log.Infof("clearing subsumed descriptor: key=%s desc=%s", k, v.(*roachpb.RangeDescriptor))
}
rdc.rangeCache.Del(k.(rangeCacheKey))
}, rangeCacheKey(meta(desc.StartKey).Next()),
rangeCacheKey(meta(desc.EndKey)))
}
// lookupReplica looks up replica by key [range]. Lookups are done
// by consulting each store in turn via Store.LookupRange(key).
// Returns RangeID and replica on success; RangeKeyMismatch error
// if not found.
// This is only for testing usage; performance doesn't matter.
func (ls *Stores) lookupReplica(start, end roachpb.RKey) (rangeID roachpb.RangeID, replica *roachpb.ReplicaDescriptor, err error) {
ls.mu.RLock()
defer ls.mu.RUnlock()
var rng *Replica
var partialDesc *roachpb.RangeDescriptor
for _, store := range ls.storeMap {
rng = store.LookupReplica(start, end)
if rng == nil {
if tmpRng := store.LookupReplica(start, nil); tmpRng != nil {
log.Warningf("range not contained in one range: [%s,%s), but have [%s,%s)",
start, end, tmpRng.Desc().StartKey, tmpRng.Desc().EndKey)
partialDesc = tmpRng.Desc()
break
}
continue
}
if replica == nil {
rangeID = rng.RangeID
replica, err = rng.GetReplica()
if err != nil {
if _, ok := err.(*errReplicaNotInRange); !ok {
return 0, nil, err
}
}
continue
}
// Should never happen outside of tests.
return 0, nil, util.Errorf(
"range %+v exists on additional store: %+v", rng, store)
}
if replica == nil {
err = roachpb.NewRangeKeyMismatchError(start.AsRawKey(), end.AsRawKey(), partialDesc)
}
return rangeID, replica, err
}
func getNode(t *testing.T, nodes map[string]storage.RangeTreeNode, testName string, key roachpb.RKey) (storage.RangeTreeNode, bool) {
if key != nil {
if node, ok := nodes[key.String()]; !ok {
t.Errorf("%s: could not locate node with key %s", testName, key)
} else {
return node, ok
}
}
return storage.RangeTreeNode{}, false
}
// ComputeSplitKeys takes a start and end key and returns an array of keys
// at which to split the span [start, end).
// The only required splits are at each user table prefix.
func (s SystemConfig) ComputeSplitKeys(startKey, endKey roachpb.RKey) []roachpb.RKey {
testingLock.Lock()
tableSplitsDisabled := testingDisableTableSplits
testingLock.Unlock()
if tableSplitsDisabled {
return nil
}
tableStart := roachpb.RKey(keys.UserTableDataMin)
if !tableStart.Less(endKey) {
// This range is before the user tables span: no required splits.
return nil
}
startID, ok := ObjectIDForKey(startKey)
if !ok || startID <= keys.MaxReservedDescID {
// The start key is either:
// - not part of the structured data span
// - part of the system span
// In either case, start looking for splits at the first ID usable
// by the user data span.
startID = keys.MaxReservedDescID + 1
} else {
// The start key is either already a split key, or after the split
// key for its ID. We can skip straight to the next one.
startID++
}
// Find the largest object ID.
// We can't keep splitting until we reach endKey as it could be roachpb.KeyMax.
endID, err := s.GetLargestObjectID()
if err != nil {
log.Errorf("unable to determine largest object ID from system config: %s", err)
return nil
}
// Build key prefixes for sequential table IDs until we reach endKey.
var splitKeys []roachpb.RKey
var key roachpb.RKey
// endID could be smaller than startID if we don't have user tables.
for id := startID; id <= endID; id++ {
key = keys.MakeTablePrefix(id)
// Skip if the range starts on a split key.
if !startKey.Less(key) {
continue
}
// Handle the case where EndKey is already a table prefix.
if !key.Less(endKey) {
break
}
splitKeys = append(splitKeys, key)
}
return splitKeys
}
// ObjectIDForKey returns the object ID (table or database) for 'key',
// or (_, false) if not within the structured key space.
func ObjectIDForKey(key roachpb.RKey) (uint32, bool) {
if key.Equal(roachpb.RKeyMax) {
return 0, false
}
if encoding.PeekType(key) != encoding.Int {
// TODO(marc): this should eventually return SystemDatabaseID.
return 0, false
}
// Consume first encoded int.
_, id64, err := encoding.DecodeUvarint(key)
return uint32(id64), err == nil
}
// LookupReplica looks up replica by key [range]. Lookups are done
// by consulting each store in turn via Store.LookupReplica(key).
// Returns RangeID and replica on success; RangeKeyMismatch error
// if not found.
// If end is nil, a replica containing start is looked up.
// This is only for testing usage; performance doesn't matter.
func (ls *Stores) LookupReplica(
start, end roachpb.RKey,
) (roachpb.RangeID, roachpb.ReplicaDescriptor, error) {
ls.mu.RLock()
defer ls.mu.RUnlock()
var rangeID roachpb.RangeID
var repDesc roachpb.ReplicaDescriptor
var repDescFound bool
for _, store := range ls.storeMap {
replica := store.LookupReplica(start, nil)
if replica == nil {
continue
}
// Verify that the descriptor contains the entire range.
if desc := replica.Desc(); !desc.ContainsKeyRange(start, end) {
log.Warningf(context.TODO(), "range not contained in one range: [%s,%s), but have [%s,%s)",
start, end, desc.StartKey, desc.EndKey)
err := roachpb.NewRangeKeyMismatchError(start.AsRawKey(), end.AsRawKey(), desc)
return 0, roachpb.ReplicaDescriptor{}, err
}
rangeID = replica.RangeID
var err error
repDesc, err = replica.GetReplicaDescriptor()
if err != nil {
if _, ok := err.(*roachpb.RangeNotFoundError); ok {
// We are not holding a lock across this block; the replica could have
// been removed from the range (via down-replication) between the
// LookupReplica and the GetReplicaDescriptor calls. In this case just
// ignore this replica.
continue
}
return 0, roachpb.ReplicaDescriptor{}, err
}
if repDescFound {
// We already found the range; this should never happen outside of tests.
err := errors.Errorf("range %+v exists on additional store: %+v", replica, store)
return 0, roachpb.ReplicaDescriptor{}, err
}
repDescFound = true
}
if !repDescFound {
return 0, roachpb.ReplicaDescriptor{}, roachpb.NewRangeNotFoundError(0)
}
return rangeID, repDesc, nil
}
// verifyBinarySearchTree checks to ensure that all keys to the left of the root
// node are less than it, and all nodes to the right of the root node are
// greater than it. It recursively walks the tree to perform this same check.
func verifyBinarySearchTree(t *testing.T, nodes map[string]storage.RangeTreeNode, testName string, node storage.RangeTreeNode, keyMin, keyMax roachpb.RKey) {
if !node.Key.Less(keyMax) {
t.Errorf("%s: Failed Property BST - The key %s is not less than %s.", testName, node.Key, keyMax)
}
// We need the extra check since roachpb.KeyMin is actually a range start key.
if !keyMin.Less(node.Key) && !node.Key.Equal(roachpb.RKeyMin) {
t.Errorf("%s: Failed Property BST - The key %s is not greater than %s.", testName, node.Key, keyMin)
}
if left, ok := getNode(t, nodes, testName, node.LeftKey); ok {
verifyBinarySearchTree(t, nodes, testName, left, keyMin, node.Key)
}
if right, ok := getNode(t, nodes, testName, node.RightKey); ok {
verifyBinarySearchTree(t, nodes, testName, right, node.Key, keyMax)
}
}
// verifyBinarySearchTree checks to ensure that all keys to the left of the root
// node are less than it, and all nodes to the right of the root node are
// greater than it. It recursively walks the tree to perform this same check.
func verifyBinarySearchTree(t *testing.T, nodes map[string]roachpb.RangeTreeNode, testName string, node *roachpb.RangeTreeNode, keyMin, keyMax roachpb.RKey) {
if node == nil {
return
}
if !node.Key.Less(keyMax) {
t.Errorf("%s: Failed Property BST - The key %s is not less than %s.", testName, node.Key, keyMax)
}
// We need the extra check since roachpb.KeyMin is actually a range start key.
if !keyMin.Less(node.Key) && !node.Key.Equal(roachpb.RKeyMin) {
t.Errorf("%s: Failed Property BST - The key %s is not greater than %s.", testName, node.Key, keyMin)
}
left, right := getLeftAndRight(t, nodes, testName, node)
verifyBinarySearchTree(t, nodes, testName, left, keyMin, node.Key)
verifyBinarySearchTree(t, nodes, testName, right, node.Key, keyMax)
}
func (db *testDescriptorDB) getDescriptors(key roachpb.RKey, considerIntents bool, useReverseScan bool) ([]roachpb.RangeDescriptor, []roachpb.RangeDescriptor, *roachpb.Error) {
rs := make([]roachpb.RangeDescriptor, 0, 1)
preRs := make([]roachpb.RangeDescriptor, 0, 2)
for i := 0; i < 3; i++ {
var endKey roachpb.RKey
if useReverseScan {
endKey = key
} else {
endKey = key.Next()
}
v := db.data.Ceil(testDescriptorNode{
&roachpb.RangeDescriptor{
EndKey: endKey,
},
})
if v == nil {
break
}
desc := *(v.(testDescriptorNode).RangeDescriptor)
if i == 0 {
rs = append(rs, desc)
if considerIntents {
desc.RangeID++
rs = append(rs, desc)
break
} else if db.disablePrefetch {
break
}
} else {
preRs = append(preRs, desc)
}
// Break to keep from skidding off the end of the available ranges.
if desc.EndKey.Equal(roachpb.RKeyMax) {
break
}
if useReverseScan {
key = desc.StartKey
} else {
key = desc.EndKey
}
}
return rs, preRs, nil
}
// RangeLookup implements the RangeDescriptorDB interface. It looks up
// the descriptors for the given (meta) key.
func (ls *Stores) RangeLookup(key roachpb.RKey, _ *roachpb.RangeDescriptor, considerIntents, useReverseScan bool) ([]roachpb.RangeDescriptor, *roachpb.Error) {
ba := roachpb.BatchRequest{}
ba.ReadConsistency = roachpb.INCONSISTENT
ba.Add(&roachpb.RangeLookupRequest{
Span: roachpb.Span{
// key is a meta key, so it's guaranteed not local-prefixed.
Key: key.AsRawKey(),
},
MaxRanges: 1,
ConsiderIntents: considerIntents,
Reverse: useReverseScan,
})
br, pErr := ls.Send(context.Background(), ba)
if pErr != nil {
return nil, pErr
}
return br.Responses[0].GetInner().(*roachpb.RangeLookupResponse).Ranges, nil
}
// ObjectIDForKey returns the object ID (table or database) for 'key',
// or (_, false) if not within the structured key space.
func ObjectIDForKey(key roachpb.RKey) (uint32, bool) {
if key.Equal(roachpb.RKeyMax) {
return 0, false
}
if key.Equal(keys.TableDataPrefix) {
// TODO(marc): this should eventually return SystemDatabaseID.
return 0, false
}
remaining := bytes.TrimPrefix(key, keys.TableDataPrefix)
if len(remaining) == len(key) {
// TrimPrefix returns the input untouched if the prefix doesn't match.
return 0, false
}
// Consume first encoded int.
_, id64, err := encoding.DecodeUvarint(remaining)
return uint32(id64), err == nil
}
// validateRangeMetaKey validates that the given key is a valid Range Metadata
// key. This checks only the constraints common to forward and backwards scans:
// correct prefix and not exceeding KeyMax.
func validateRangeMetaKey(key roachpb.RKey) error {
// KeyMin is a valid key.
if key.Equal(roachpb.RKeyMin) {
return nil
}
// Key must be at least as long as Meta1Prefix.
if len(key) < len(Meta1Prefix) {
return NewInvalidRangeMetaKeyError("too short", key)
}
prefix, body := key[:len(Meta1Prefix)], key[len(Meta1Prefix):]
if !prefix.Equal(Meta2Prefix) && !prefix.Equal(Meta1Prefix) {
return NewInvalidRangeMetaKeyError("not a meta key", key)
}
if roachpb.RKeyMax.Less(body) {
return NewInvalidRangeMetaKeyError("body of meta key range lookup is > KeyMax", key)
}
return nil
}
// 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 roachpb.RKey, inclusive bool) (
rangeCacheKey, *roachpb.RangeDescriptor) {
// The cache is indexed using the end-key of the range, but the
// end-key is non-inclusive by default.
var metaKey roachpb.RKey
if !inclusive {
metaKey = meta(key.Next())
} else {
metaKey = meta(key)
}
k, v, ok := rdc.rangeCache.Ceil(rangeCacheKey(metaKey))
if !ok {
return nil, nil
}
metaEndKey := k.(rangeCacheKey)
rd := v.(*roachpb.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
}
// MetaReverseScanBounds returns the range [start,end) within which the desired
// meta record can be found by means of a reverse engine scan. The given key
// must be a valid RangeMetaKey as defined by validateRangeMetaKey.
func MetaReverseScanBounds(key roachpb.RKey) (roachpb.Key, roachpb.Key, error) {
if err := validateRangeMetaKey(key); err != nil {
return nil, nil, err
}
if key.Equal(roachpb.RKeyMin) || key.Equal(Meta1Prefix) {
return nil, nil, NewInvalidRangeMetaKeyError("KeyMin and Meta1Prefix can't be used as the key of reverse scan", key)
}
if key.Equal(Meta2Prefix) {
// Special case Meta2Prefix: this is the first key in Meta2, and the scan
// interval covers all of Meta1.
return Meta1Prefix, key.Next().AsRawKey(), nil
}
// Otherwise find the first entry greater than the given key and find the last entry
// in the same prefix. For MVCCReverseScan the endKey is exclusive, if we want to find
// the range descriptor the given key specified,we need to set the key.Next() as the
// MVCCReverseScan`s endKey. For example:
// If we have ranges [a,f) and [f,z), then we'll have corresponding meta records
// at f and z. If you're looking for the meta record for key f, then you want the
// second record (exclusive in MVCCReverseScan), hence key.Next() below.
return key[:len(Meta1Prefix)].AsRawKey(), key.Next().AsRawKey(), nil
}
// LookupReplica looks up replica by key [range]. Lookups are done
// by consulting each store in turn via Store.LookupReplica(key).
// Returns RangeID and replica on success; RangeKeyMismatch error
// if not found.
// If end is nil, a replica containing start is looked up.
// This is only for testing usage; performance doesn't matter.
func (ls *Stores) LookupReplica(start, end roachpb.RKey) (rangeID roachpb.RangeID, repDesc roachpb.ReplicaDescriptor, err error) {
ls.mu.RLock()
defer ls.mu.RUnlock()
var rng *Replica
var partialDesc *roachpb.RangeDescriptor
var repDescFound bool
for _, store := range ls.storeMap {
rng = store.LookupReplica(start, end)
if rng == nil {
if tmpRng := store.LookupReplica(start, nil); tmpRng != nil {
partialDesc = tmpRng.Desc()
log.Warningf(context.TODO(), "range not contained in one range: [%s,%s), but have [%s,%s)",
start, end, partialDesc.StartKey, partialDesc.EndKey)
break
}
continue
}
if !repDescFound {
rangeID = rng.RangeID
repDesc, err = rng.GetReplicaDescriptor()
if err != nil {
if _, ok := err.(*roachpb.RangeNotFoundError); !ok {
return 0, roachpb.ReplicaDescriptor{}, err
}
} else {
repDescFound = true
}
continue
}
// Should never happen outside of tests.
return 0, roachpb.ReplicaDescriptor{}, errors.Errorf(
"range %+v exists on additional store: %+v", rng, store,
)
}
if !repDescFound {
err = roachpb.NewRangeKeyMismatchError(start.AsRawKey(), end.AsRawKey(), partialDesc)
}
return rangeID, repDesc, err
}
// verifyBinarySearchTree checks to ensure that all keys to the left of the root
// node are less than it, and all nodes to the right of the root node are
// greater than it. It recursively walks the tree to perform this same check.
func verifyBinarySearchTree(t *testing.T, tc *treeContext, testName string, node *roachpb.RangeTreeNode, keyMin, keyMax roachpb.RKey) {
if !node.Key.Less(keyMax) {
t.Errorf("%s: Failed Property BST - The key %s is not less than %s.", testName, node.Key, keyMax)
}
if !keyMin.Less(node.Key) {
t.Errorf("%s: Failed Property BST - The key %s is not greater than %s.", testName, node.Key, keyMin)
}
if node.LeftKey != nil {
left, err := tc.getNode(node.LeftKey)
if err != nil {
t.Fatal(err)
}
verifyBinarySearchTree(t, tc, testName, left, keyMin, node.Key)
}
if node.RightKey != nil {
right, err := tc.getNode(node.RightKey)
if err != nil {
t.Fatal(err)
}
verifyBinarySearchTree(t, tc, testName, right, node.Key, keyMax)
}
}
// MetaScanBounds returns the range [start,end) within which the desired meta
// record can be found by means of an engine scan. The given key must be a
// valid RangeMetaKey as defined by validateRangeMetaKey.
// TODO(tschottdorf): a lot of casting going on inside.
func MetaScanBounds(key roachpb.RKey) (roachpb.Key, roachpb.Key, error) {
if err := validateRangeMetaKey(key); err != nil {
return nil, nil, err
}
if key.Equal(Meta2KeyMax) {
return nil, nil, NewInvalidRangeMetaKeyError("Meta2KeyMax can't be used as the key of scan", key)
}
if key.Equal(roachpb.RKeyMin) {
// Special case KeyMin: find the first entry in meta1.
return Meta1Prefix, Meta1Prefix.PrefixEnd(), nil
}
if key.Equal(Meta1KeyMax) {
// Special case Meta1KeyMax: this is the last key in Meta1, we don't want
// to start at Next().
return Meta1KeyMax, Meta1Prefix.PrefixEnd(), nil
}
// Otherwise find the first entry greater than the given key in the same meta prefix.
return key.Next().AsRawKey(), key[:len(Meta1Prefix)].PrefixEnd().AsRawKey(), nil
}
// truncate restricts all contained requests to the given key range.
// Even on error, the returned closure must be executed; it undoes any
// truncations performed.
// First, the boundaries of the truncation are obtained: This is the
// intersection between [from,to) and the descriptor's range.
// Secondly, all requests contained in the batch are "truncated" to
// the resulting range, inserting NoopRequest appropriately to
// replace requests which are left without a key range to operate on.
// The number of non-noop requests after truncation is returned along
// with a closure which must be executed to undo the truncation, even
// in case of an error.
// TODO(tschottdorf): Consider returning a new BatchRequest, which has more
// overhead in the common case of a batch which never needs truncation but is
// less magical.
func truncate(br *roachpb.BatchRequest, desc *roachpb.RangeDescriptor, from, to roachpb.RKey) (func(), int, error) {
if !desc.ContainsKey(from) {
from = desc.StartKey
}
if !desc.ContainsKeyRange(desc.StartKey, to) || to == nil {
to = desc.EndKey
}
truncateOne := func(args roachpb.Request) (bool, []func(), error) {
if _, ok := args.(*roachpb.NoopRequest); ok {
return true, nil, nil
}
header := args.Header()
if !roachpb.IsRange(args) {
// This is a point request.
if len(header.EndKey) > 0 {
return false, nil, util.Errorf("%T is not a range command, but EndKey is set", args)
}
if !desc.ContainsKey(keys.Addr(header.Key)) {
return true, nil, nil
}
return false, nil, nil
}
// We're dealing with a range-spanning request.
var undo []func()
keyAddr, endKeyAddr := keys.Addr(header.Key), keys.Addr(header.EndKey)
if l, r := !keyAddr.Equal(header.Key), !endKeyAddr.Equal(header.EndKey); l || r {
if !desc.ContainsKeyRange(keyAddr, endKeyAddr) {
return false, nil, util.Errorf("local key range must not span ranges")
}
if !l || !r {
return false, nil, util.Errorf("local key mixed with global key in range")
}
}
// Below, {end,}keyAddr equals header.{End,}Key, so nothing is local.
if keyAddr.Less(from) {
{
origKey := header.Key
undo = append(undo, func() { header.Key = origKey })
}
header.Key = from.AsRawKey() // "from" can't be local
keyAddr = from
}
if !endKeyAddr.Less(to) {
{
origEndKey := header.EndKey
undo = append(undo, func() { header.EndKey = origEndKey })
}
header.EndKey = to.AsRawKey() // "to" can't be local
endKeyAddr = to
}
// Check whether the truncation has left any keys in the range. If not,
// we need to cut it out of the request.
return !keyAddr.Less(endKeyAddr), undo, nil
}
var fns []func()
gUndo := func() {
for _, f := range fns {
f()
}
}
var numNoop int
for pos, arg := range br.Requests {
omit, undo, err := truncateOne(arg.GetInner())
if omit {
numNoop++
nReq := &roachpb.RequestUnion{}
if !nReq.SetValue(&roachpb.NoopRequest{}) {
panic("RequestUnion excludes NoopRequest")
}
oReq := br.Requests[pos]
br.Requests[pos] = *nReq
posCpy := pos // for closure
undo = append(undo, func() {
br.Requests[posCpy] = oReq
})
}
fns = append(fns, undo...)
if err != nil {
return gUndo, 0, err
}
}
return gUndo, len(br.Requests) - numNoop, nil
}
// ComputeSplitKeys takes a start and end key and returns an array of keys
// at which to split the span [start, end).
// The only required splits are at each user table prefix.
func (s SystemConfig) ComputeSplitKeys(startKey, endKey roachpb.RKey) []roachpb.RKey {
if TestingTableSplitsDisabled() {
return nil
}
tableStart := roachpb.RKey(keys.ReservedTableDataMin)
if !tableStart.Less(endKey) {
// This range is before the user tables span: no required splits.
return nil
}
startID, ok := ObjectIDForKey(startKey)
if !ok || startID <= keys.MaxSystemDescID {
// The start key is either:
// - not part of the structured data span
// - part of the system span
// In either case, start looking for splits at the first ID usable
// by the user data span.
startID = keys.MaxSystemDescID + 1
} else {
// The start key is either already a split key, or after the split
// key for its ID. We can skip straight to the next one.
startID++
}
// Build key prefixes for sequential table IDs until we reach endKey. Note
// that there are two disjoint sets of sequential keys: non-system reserved
// tables have sequential IDs, as do user tables, but the two ranges contain a
// gap.
var splitKeys []roachpb.RKey
var key roachpb.RKey
// appendSplitKeys generates all possible split keys between the given range
// of IDs and adds them to splitKeys.
appendSplitKeys := func(startID, endID uint32) {
// endID could be smaller than startID if we don't have user tables.
for id := startID; id <= endID; id++ {
key = keys.MakeNonColumnKey(keys.MakeTablePrefix(id))
// Skip if this ID matches the startKey passed to ComputeSplitKeys.
if !startKey.Less(key) {
continue
}
// Handle the case where EndKey is already a table prefix.
if !key.Less(endKey) {
break
}
splitKeys = append(splitKeys, key)
}
}
// If the startKey falls within the non-system reserved range, compute those
// keys first.
if startID <= keys.MaxReservedDescID {
endID, err := s.GetLargestObjectID(keys.MaxReservedDescID)
if err != nil {
log.Errorf("unable to determine largest reserved object ID from system config: %s", err)
return nil
}
appendSplitKeys(startID, endID)
startID = keys.MaxReservedDescID + 1
}
// Append keys in the user space.
endID, err := s.GetLargestObjectID(0)
if err != nil {
log.Errorf("unable to determine largest object ID from system config: %s", err)
return nil
}
appendSplitKeys(startID, endID)
return splitKeys
}