// 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.Key, options lookupOptions) (*roachpb.RangeDescriptor, bool, func(), *roachpb.Error) {
var desc *roachpb.RangeDescriptor
var err error
var descKey roachpb.Key
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
}
// Add the specified timestamp to the cache as covering the range of
// keys from start to end. If end is nil, the range covers the start
// key only. txnID is nil for no transaction. readOnly specifies
// whether the command adding this timestamp was read-only or not.
func (tc *TimestampCache) Add(start, end roachpb.Key, timestamp roachpb.Timestamp, txnID []byte, readOnly bool) {
// This gives us a memory-efficient end key if end is empty.
if len(end) == 0 {
end = start.Next()
start = end[:len(start)]
}
if tc.latest.Less(timestamp) {
tc.latest = timestamp
}
// Only add to the cache if the timestamp is more recent than the
// low water mark.
if tc.lowWater.Less(timestamp) {
// Check existing, overlapping entries. Remove superseded
// entries or return without adding this entry if necessary.
key := tc.cache.NewKey(start, end)
for _, o := range tc.cache.GetOverlaps(start, end) {
ce := o.Value.(cacheEntry)
if ce.readOnly != readOnly {
continue
}
if o.Key.Contains(key) && !ce.timestamp.Less(timestamp) {
return // don't add this key; there's already a cache entry with >= timestamp.
} else if key.Contains(o.Key) && !timestamp.Less(ce.timestamp) {
tc.cache.Del(o.Key) // delete existing key; this cache entry supersedes.
}
}
ce := cacheEntry{timestamp: timestamp, txnID: txnID, readOnly: readOnly}
tc.cache.Add(key, ce)
}
}
// clearOverlappingCachedRangeDescriptors looks up and clears any
// cache entries which overlap the specified key or descriptor.
func (rdc *rangeDescriptorCache) clearOverlappingCachedRangeDescriptors(key, metaKey roachpb.Key, 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(keys.RangeMetaKey(desc.StartKey).Next()),
rangeCacheKey(keys.RangeMetaKey(desc.EndKey)))
}
// 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 'proto'.
func prev(ba roachpb.BatchRequest, k roachpb.Key) roachpb.Key {
candidate := roachpb.KeyMin
for _, union := range ba.Requests {
h := union.GetInner().Header()
addr := keys.KeyAddress(h.Key)
eAddr := keys.KeyAddress(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
}
// prettyPrintInternal parse key with prefix in keyDict,
// if the key don't march any prefix in keyDict, return its byte value with quotation and false,
// or else return its human readable value and true.
func prettyPrintInternal(key roachpb.Key) (string, bool) {
var buf bytes.Buffer
for _, k := range keyDict {
if key.Compare(k.start) >= 0 && (k.end == nil || key.Compare(k.end) <= 0) {
buf.WriteString(k.name)
if k.end != nil && k.end.Compare(key) == 0 {
buf.WriteString("/Max")
return buf.String(), true
}
hasPrefix := false
for _, e := range k.entries {
if bytes.HasPrefix(key, e.prefix) {
hasPrefix = true
key = key[len(e.prefix):]
fmt.Fprintf(&buf, "%s%s", e.name, e.ppFunc(key))
break
}
}
if !hasPrefix {
key = key[len(k.start):]
fmt.Fprintf(&buf, "/%q", []byte(key))
}
return buf.String(), true
}
}
return fmt.Sprintf("%q", []byte(key)), false
}
// addKeyRange adds the specified key range to the interval cache,
// taking care not to add this range if existing entries already
// completely cover the range.
func (tm *txnMetadata) addKeyRange(start, end roachpb.Key) {
// This gives us a memory-efficient end key if end is empty.
// The most common case for keys in the intents interval map
// is for single keys. However, the interval cache requires
// a non-empty interval, so we create two key slices which
// share the same underlying byte array.
if len(end) == 0 {
end = start.Next()
start = end[:len(start)]
}
key := tm.keys.MakeKey(start, end)
for _, o := range tm.keys.GetOverlaps(key.Start, key.End) {
if o.Key.Contains(key) {
return
} else if key.Contains(*o.Key) {
tm.keys.Del(o.Key)
}
}
// Since no existing key range fully covered this range, add it now. The
// strange assignment to pkey makes sure we delay the heap allocation until
// we know it is necessary.
alloc := struct {
key cache.IntervalKey
entry cache.Entry
}{key: key}
alloc.entry.Key = &alloc.key
tm.keys.AddEntry(&alloc.entry)
}
// Encodes datum at the end of key, using direction `dir` for the encoding.
// The key is a span end key, which is exclusive, but `val` needs to
// be inclusive. So if datum is the last end constraint, we transform it accordingly.
func encodeInclusiveEndValue(
key roachpb.Key, datum parser.Datum, dir encoding.Direction,
isLastEndConstraint bool) roachpb.Key {
// Since the end of a span is exclusive, if the last constraint is an
// inclusive one, we might need to make the key exclusive by applying a
// PrefixEnd(). We normally avoid doing this by transforming "a = x" to
// "a = x±1" for the last end constraint, depending on the encoding direction
// (since this keeps the key nice and pretty-printable).
// However, we might not be able to do the ±1.
needExclusiveKey := false
if isLastEndConstraint {
if dir == encoding.Ascending {
if datum.IsMax() {
needExclusiveKey = true
} else {
datum = datum.Next()
}
} else {
if datum.IsMin() || !datum.HasPrev() {
needExclusiveKey = true
} else {
datum = datum.Prev()
}
}
}
key, pErr := encodeTableKey(key, datum, dir)
if pErr != nil {
panic(pErr)
}
if needExclusiveKey {
key = key.PrefixEnd()
}
return key
}
// GetIndex searches the kv list for 'key' and returns its index if found.
func (s SystemConfig) GetIndex(key roachpb.Key) (int, bool) {
l := len(s.Values)
index := sort.Search(l, func(i int) bool {
return bytes.Compare(s.Values[i].Key, key) >= 0
})
if index == l || !key.Equal(s.Values[index].Key) {
return 0, false
}
return index, true
}
// prettyKey pretty-prints the specified key, skipping over the first `skip`
// fields. The pretty printed key looks like:
//
// /Table/<tableID>/<indexID>/...
//
// We always strip off the /Table prefix and then `skip` more fields. Note that
// this assumes that the fields themselves do not contain '/', but that is
// currently true for the fields we care about stripping (the table and index
// ID).
func prettyKey(key roachpb.Key, skip int) string {
p := key.String()
for i := 0; i <= skip; i++ {
n := strings.IndexByte(p[1:], '/')
if n == -1 {
return ""
}
p = p[n+1:]
}
return p
}
// 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.Key) []roachpb.Key {
if TestingDisableTableSplits {
return nil
}
tableStart := roachpb.Key(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.KeySlice
var key roachpb.Key
// 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
}
// GetIndex searches the kv list for 'key' and returns its index if found.
func (s *SystemConfig) GetIndex(key roachpb.Key) (int, bool) {
if s == nil {
return 0, false
}
l := len(s.Values)
index := sort.Search(l, func(i int) bool {
return !s.Values[i].Key.Less(key)
})
if index == l || !key.Equal(s.Values[index].Key) {
return 0, false
}
return index, true
}
// 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.Key) {
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.KeyMin) {
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)
}
// addKeyRange adds the specified key range to the range group,
// taking care not to add this range if existing entries already
// completely cover the range.
func addKeyRange(keys interval.RangeGroup, start, end roachpb.Key) {
// This gives us a memory-efficient end key if end is empty.
// The most common case for keys in the intents interval map
// is for single keys. However, the range group requires
// a non-empty interval, so we create two key slices which
// share the same underlying byte array.
if len(end) == 0 {
end = start.Next()
start = end[:len(start)]
}
keyR := interval.Range{
Start: interval.Comparable(start),
End: interval.Comparable(end),
}
keys.Add(keyR)
}
// TODO(dt): Batch checks of many rows.
func (f baseFKHelper) check(values parser.DTuple) (parser.DTuple, error) {
var key roachpb.Key
if values != nil {
keyBytes, _, err := sqlbase.EncodeIndexKey(f.searchIdx, f.ids, values, f.searchPrefix)
if err != nil {
return nil, err
}
key = roachpb.Key(keyBytes)
} else {
key = roachpb.Key(f.searchPrefix)
}
spans := sqlbase.Spans{sqlbase.Span{Start: key, End: key.PrefixEnd()}}
if err := f.rf.StartScan(f.txn, spans, 1); err != nil {
return nil, err
}
return f.rf.NextRow()
}
// GetMax returns the maximum read and write timestamps which overlap
// the interval spanning from start to end. Cached timestamps matching
// the specified txnID are not considered. If no part of the specified
// range is overlapped by timestamps in the cache, the low water
// timestamp is returned for both read and write timestamps.
//
// The txn ID prevents restarts with a pattern like: read("a"),
// write("a"). The read adds a timestamp for "a". Then the write (for
// the same transaction) would get that as the max timestamp and be
// forced to increment it. This allows timestamps from the same txn
// to be ignored.
func (tc *TimestampCache) GetMax(start, end roachpb.Key, txnID []byte) (roachpb.Timestamp, roachpb.Timestamp) {
if len(end) == 0 {
end = start.Next()
}
maxR := tc.lowWater
maxW := tc.lowWater
for _, o := range tc.cache.GetOverlaps(start, end) {
ce := o.Value.(*cacheValue)
if ce.txnID == nil || txnID == nil || !roachpb.TxnIDEqual(txnID, ce.txnID) {
if ce.readOnly && maxR.Less(ce.timestamp) {
maxR = ce.timestamp
} else if !ce.readOnly && maxW.Less(ce.timestamp) {
maxW = ce.timestamp
}
}
}
return maxR, maxW
}
// ObjectIDForKey returns the object ID (table or database) for 'key',
// or (_, false) if not within the structured key space.
func ObjectIDForKey(key roachpb.Key) (uint32, bool) {
if key.Equal(roachpb.KeyMax) {
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
}
func (tc *TimestampCache) getMax(start, end roachpb.Key, txnID *uuid.UUID, readTSCache bool) roachpb.Timestamp {
if len(end) == 0 {
end = start.ShallowNext()
}
max := tc.lowWater
cache := tc.wCache
if readTSCache {
cache = tc.rCache
}
for _, o := range cache.GetOverlaps(start, end) {
ce := o.Value.(*cacheValue)
if ce.txnID == nil || txnID == nil || !roachpb.TxnIDEqual(txnID, ce.txnID) {
if max.Less(ce.timestamp) {
max = ce.timestamp
}
}
}
return max
}
// 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.Key) roachpb.Key {
candidate := roachpb.KeyMax
for _, union := range ba.Requests {
h := union.GetInner().Header()
addr := keys.KeyAddress(h.Key)
if addr.Less(k) {
if eAddr := keys.KeyAddress(h.EndKey); k.Less(eAddr) {
// Starts below k, but continues beyond. Need to stay at k.
return k
}
// Affects only [KeyMin,k).
continue
}
// We want the smallest of the surviving candidates.
if addr.Less(candidate) {
candidate = addr
}
}
return candidate
}
// PrettyPrint prints the key in a human readable format:
//
// Key's Format Key's Value
// /Local/... "\x01"+...
// /Store/... "\x01s"+...
// /RangeID/... "\x01s"+[rangeid]
// /[rangeid]/AbortCache/[id] "\x01s"+[rangeid]+"abc-"+[id]
// /[rangeid]/RaftLeaderLease "\x01s"+[rangeid]+"rfll"
// /[rangeid]/RaftTombstone "\x01s"+[rangeid]+"rftb"
// /[rangeid]/RaftHardState "\x01s"+[rangeid]+"rfth"
// /[rangeid]/RaftAppliedIndex "\x01s"+[rangeid]+"rfta"
// /[rangeid]/RaftLog/logIndex:[logIndex] "\x01s"+[rangeid]+"rftl"+[logIndex]
// /[rangeid]/RaftTruncatedState "\x01s"+[rangeid]+"rftt"
// /[rangeid]/RaftLastIndex "\x01s"+[rangeid]+"rfti"
// /[rangeid]/RangeLastReplicaGCTimestamp "\x01s"+[rangeid]+"rlrt"
// /[rangeid]/RangeLastVerificationTimestamp "\x01s"+[rangeid]+"rlvt"
// /[rangeid]/RangeStats "\x01s"+[rangeid]+"stat"
// /Range/... "\x01k"+...
// /RangeDescriptor/[key] "\x01k"+[key]+"rdsc"
// /RangeTreeNode/[key] "\x01k"+[key]+"rtn-"
// /Transaction/addrKey:[key]/id:[id] "\x01k"+[key]+"txn-"+[id]
// /Local/Max "\x02"
//
// /Meta1/[key] "\x02"+[key]
// /Meta2/[key] "\x03"+[key]
// /System/... "\x04"
// /StatusNode/[key] "\x04status-node-"+[key]
// /System/Max "\x05"
//
// /Table/[key] [key]
//
// /Min ""
// /Max "\xff\xff"
func PrettyPrint(key roachpb.Key) string {
for _, k := range constKeyDict {
if key.Equal(k.value) {
return k.name
}
}
for _, k := range keyOfKeyDict {
if bytes.HasPrefix(key, k.prefix) {
key = key[len(k.prefix):]
str, formatted := prettyPrintInternal(key)
if formatted {
return k.name + str
}
return k.name + "/" + str
}
}
str, _ := prettyPrintInternal(key)
return str
}
// 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.Key) error {
// KeyMin is a valid key.
if key.Equal(roachpb.KeyMin) {
return nil
}
// Key must be at least as long as Meta1Prefix.
if len(key) < len(Meta1Prefix) {
return NewInvalidRangeMetaKeyError("too short", key)
}
prefix, body := roachpb.Key(key[:len(Meta1Prefix)]), roachpb.Key(key[len(Meta1Prefix):])
if !prefix.Equal(Meta2Prefix) && !prefix.Equal(Meta1Prefix) {
return NewInvalidRangeMetaKeyError("not a meta key", key)
}
if roachpb.KeyMax.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.Key, 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.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.(*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.Key) (roachpb.Key, roachpb.Key, error) {
if err := validateRangeMetaKey(key); err != nil {
return nil, nil, err
}
if key.Equal(roachpb.KeyMin) || 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(), 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)], key.Next(), nil
}
// addKeyRange adds the specified key range to the interval cache,
// taking care not to add this range if existing entries already
// completely cover the range.
func (tm *txnMetadata) addKeyRange(start, end roachpb.Key) {
// This gives us a memory-efficient end key if end is empty.
// The most common case for keys in the intents interval map
// is for single keys. However, the interval cache requires
// a non-empty interval, so we create two key slices which
// share the same underlying byte array.
if len(end) == 0 {
end = start.Next()
start = end[:len(start)]
}
key := tm.keys.NewKey(start, end)
for _, o := range tm.keys.GetOverlaps(start, end) {
if o.Key.Contains(key) {
return
} else if key.Contains(o.Key) {
tm.keys.Del(o.Key)
}
}
// Since no existing key range fully covered this range, add it now.
tm.keys.Add(key, 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, tc *treeContext, testName string, node *roachpb.RangeTreeNode, keyMin, keyMax roachpb.Key) {
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)
}
}
// PrettyPrint prints the key in a human readable format:
//
// Key's Format Key's Value
// /Local/... "\x01"+...
// /Store/... "\x01s"+...
// /RangeID/... "\x01s"+[rangeid]
// /[rangeid]/SequenceCache/[id]/seq:[seq] "\x01s"+[rangeid]+"res-"+[id]+[seq]
// /[rangeid]/RaftLeaderLease "\x01s"+[rangeid]+"rfll"
// /[rangeid]/RaftTombstone "\x01s"+[rangeid]+"rftb"
// /[rangeid]/RaftHardState "\x01s"+[rangeid]+"rfth"
// /[rangeid]/RaftAppliedIndex "\x01s"+[rangeid]+"rfta"
// /[rangeid]/RaftLog/logIndex:[logIndex] "\x01s"+[rangeid]+"rftl"+[logIndex]
// /[rangeid]/RaftTruncatedState "\x01s"+[rangeid]+"rftt"
// /[rangeid]/RaftLastIndex "\x01s"+[rangeid]+"rfti"
// /[rangeid]/RangeLastVerificationTimestamp "\x01s"+[rangeid]+"rlvt"
// /[rangeid]/RangeStats "\x01s"+[rangeid]+"stat"
// /Range/... "\x01k"+...
// /RangeDescriptor/[key] "\x01k"+[key]+"rdsc"
// /RangeTreeNode/[key] "\x01k"+[key]+"rtn-"
// /Transaction/addrKey:[key]/id:[id] "\x01k"+[key]+"txn-"+[id]
// /Local/Max "\x02"
//
// /Meta1/[key] "\x02"+[key]
// /Meta2/[key] "\x03"+[key]
// /System/... "\x04"
// /StatusStore/[key] "\x04status-store-"+[key]
// /StatusNode/[key] "\x04status-node-"+[key]
// /System/Max "\x05"
//
// /Table/[key] [key]
//
// /Min ""
// /Max "\xff\xff"
func PrettyPrint(key roachpb.Key) string {
if bytes.Equal(key, MaxKey) {
return "/Max"
} else if bytes.Equal(key, MinKey) {
return "/Min"
}
var buf bytes.Buffer
for _, k := range keyDict {
if key.Compare(k.start) >= 0 && (k.end == nil || key.Compare(k.end) <= 0) {
fmt.Fprintf(&buf, "%s", k.name)
if k.end != nil && k.end.Compare(key) == 0 {
fmt.Fprintf(&buf, "/Max")
return buf.String()
}
hasPrefix := false
for _, e := range k.entries {
if bytes.HasPrefix(key, e.prefix) {
hasPrefix = true
key = key[len(e.prefix):]
fmt.Fprintf(&buf, "%s%s", e.name, e.ppFunc(key))
break
}
}
if !hasPrefix {
key = key[len(k.start):]
fmt.Fprintf(&buf, "/%q", []byte(key))
}
return buf.String()
}
}
return fmt.Sprintf("%q", []byte(key))
}
// 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.
func MetaScanBounds(key roachpb.Key) (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.KeyMin) {
// 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 key, Meta1Prefix.PrefixEnd(), nil
}
// Otherwise find the first entry greater than the given key in the same meta prefix.
return key.Next(), roachpb.Key(key[:len(Meta1Prefix)]).PrefixEnd(), nil
}
func (sc *SchemaChanger) truncateAndBackfillColumnsChunk(
added []sqlbase.ColumnDescriptor,
dropped []sqlbase.ColumnDescriptor,
defaultExprs []parser.TypedExpr,
evalCtx *parser.EvalContext,
sp sqlbase.Span,
) (roachpb.Key, bool, error) {
var curIndexKey roachpb.Key
done := false
err := sc.db.Txn(func(txn *client.Txn) error {
tableDesc, err := getTableDescFromID(txn, sc.tableID)
if err != nil {
return err
}
// Short circuit the backfill if the table has been deleted.
if tableDesc.Deleted() {
done = true
return nil
}
updateCols := append(added, dropped...)
fkTables := TablesNeededForFKs(*tableDesc, CheckUpdates)
for k := range fkTables {
if fkTables[k], err = getTableDescFromID(txn, k); err != nil {
return err
}
}
// TODO(dan): Tighten up the bound on the requestedCols parameter to
// makeRowUpdater.
requestedCols := make([]sqlbase.ColumnDescriptor, 0, len(tableDesc.Columns)+len(added))
requestedCols = append(requestedCols, tableDesc.Columns...)
requestedCols = append(requestedCols, added...)
ru, err := makeRowUpdater(
txn, tableDesc, fkTables, updateCols, requestedCols, rowUpdaterOnlyColumns,
)
if err != nil {
return err
}
// TODO(dan): This check is an unfortunate bleeding of the internals of
// rowUpdater. Extract the sql row to k/v mapping logic out into something
// usable here.
if !ru.isColumnOnlyUpdate() {
panic("only column data should be modified, but the rowUpdater is configured otherwise")
}
// Run a scan across the table using the primary key. Running
// the scan and applying the changes in many transactions is
// fine because the schema change is in the correct state to
// handle intermediate OLTP commands which delete and add
// values during the scan.
var rf sqlbase.RowFetcher
colIDtoRowIndex := colIDtoRowIndexFromCols(tableDesc.Columns)
valNeededForCol := make([]bool, len(tableDesc.Columns))
for i := range valNeededForCol {
_, valNeededForCol[i] = ru.fetchColIDtoRowIndex[tableDesc.Columns[i].ID]
}
err = rf.Init(tableDesc, colIDtoRowIndex, &tableDesc.PrimaryIndex, false, false,
tableDesc.Columns, valNeededForCol)
if err != nil {
return err
}
// StartScan uses 0 as a sentinal for the default limit of entries scanned.
if err := rf.StartScan(txn, sqlbase.Spans{sp}, 0); err != nil {
return err
}
indexKeyPrefix := sqlbase.MakeIndexKeyPrefix(tableDesc, tableDesc.PrimaryIndex.ID)
oldValues := make(parser.DTuple, len(ru.fetchCols))
updateValues := make(parser.DTuple, len(updateCols))
writeBatch := &client.Batch{}
var i int
for ; i < ColumnTruncateAndBackfillChunkSize; i++ {
row, err := rf.NextRow()
if err != nil {
return err
}
if row == nil {
break // Done
}
curIndexKey, _, err = sqlbase.EncodeIndexKey(
tableDesc, &tableDesc.PrimaryIndex, colIDtoRowIndex, row, indexKeyPrefix)
for j, col := range added {
if defaultExprs == nil || defaultExprs[j] == nil {
updateValues[j] = parser.DNull
} else {
updateValues[j], err = defaultExprs[j].Eval(evalCtx)
if err != nil {
return err
}
}
if !col.Nullable && updateValues[j].Compare(parser.DNull) == 0 {
return sqlbase.NewNonNullViolationError(col.Name)
}
}
for j := range dropped {
updateValues[j+len(added)] = parser.DNull
}
copy(oldValues, row)
for j := len(row); j < len(oldValues); j++ {
oldValues[j] = parser.DNull
}
if _, err := ru.updateRow(writeBatch, oldValues, updateValues); err != nil {
return err
}
}
if i < ColumnTruncateAndBackfillChunkSize {
done = true
}
if err := txn.Run(writeBatch); err != nil {
return convertBackfillError(tableDesc, writeBatch)
}
return nil
})
return curIndexKey.PrefixEnd(), done, err
}
// fetch retrieves spans from the kv
func (f *kvFetcher) fetch() error {
batchSize := f.getBatchSize()
b := &client.Batch{}
b.Header.MaxScanResults = batchSize
var resumeKey roachpb.Key
if len(f.kvs) > 0 {
resumeKey = f.kvs[len(f.kvs)-1].Key
// To resume forward scans we will set the (inclusive) scan start to the Next of the last
// received key. To resume reverse scans we will set the (exclusive) scan end to the last
// received key.
if !f.reverse {
resumeKey = resumeKey.Next()
}
}
atEnd := true
if !f.reverse {
for i := 0; i < len(f.spans); i++ {
start := f.spans[i].Start
if resumeKey != nil {
if resumeKey.Compare(f.spans[i].End) >= 0 {
// We are resuming from a key after this span.
continue
}
if resumeKey.Compare(start) > 0 {
// We are resuming from a key inside this span.
// In this case we should technically reduce the max count for the span; but
// since this count is only an optimization it's not incorrect to retrieve more
// keys for the span.
start = resumeKey
}
}
atEnd = false
b.Scan(start, f.spans[i].End, f.spans[i].Count)
}
} else {
for i := len(f.spans) - 1; i >= 0; i-- {
end := f.spans[i].End
if resumeKey != nil {
if resumeKey.Compare(f.spans[i].Start) <= 0 {
// We are resuming from a key before this span.
continue
}
if resumeKey.Compare(end) < 0 {
// We are resuming from a key inside this span.
// In this case we should technically reduce the max count for the span; but
// since this count is only an optimization it's not incorrect to retrieve more
// keys for the span.
end = resumeKey
}
}
atEnd = false
b.ReverseScan(f.spans[i].Start, end, f.spans[i].Count)
}
}
if atEnd {
// The last scan happened to finish just at the end of the last span.
f.kvs = nil
f.fetchEnd = true
return nil
}
if err := f.txn.Run(b); err != nil {
return err
}
if f.kvs == nil {
numResults := 0
for _, result := range b.Results {
numResults += len(result.Rows)
}
f.kvs = make([]client.KeyValue, 0, numResults)
} else {
f.kvs = f.kvs[:0]
}
for _, result := range b.Results {
f.kvs = append(f.kvs, result.Rows...)
}
f.batchIdx++
f.totalFetched += int64(len(f.kvs))
f.kvIndex = 0
if int64(len(f.kvs)) < batchSize {
f.fetchEnd = true
}
// TODO(radu): We should fetch the next chunk in the background instead of waiting for the next
// call to fetch(). We can use a pool of workers to issue the KV ops which will also limit the
// total number of fetches that happen in parallel (and thus the amount of resources we use).
return nil
}
func (sc *SchemaChanger) truncateAndBackfillColumnsChunk(
added []sqlbase.ColumnDescriptor,
dropped []sqlbase.ColumnDescriptor,
nonNullableColumn string,
defaultExprs []parser.TypedExpr,
evalCtx parser.EvalContext,
sp sqlbase.Span,
) (roachpb.Key, bool, error) {
var curSentinel roachpb.Key
done := false
err := sc.db.Txn(func(txn *client.Txn) error {
tableDesc, err := getTableDescFromID(txn, sc.tableID)
if err != nil {
return err
}
// Short circuit the backfill if the table has been deleted.
if tableDesc.Deleted {
done = true
return nil
}
// Run a scan across the table using the primary key. Running
// the scan and applying the changes in many transactions is
// fine because the schema change is in the correct state to
// handle intermediate OLTP commands which delete and add
// values during the scan.
b := &client.Batch{}
b.Scan(sp.Start, sp.End, ColumnTruncateAndBackfillChunkSize)
if err := txn.Run(b); err != nil {
return err
}
// Use a different batch to truncate/backfill columns.
writeBatch := &client.Batch{}
marshalled := make([]roachpb.Value, len(defaultExprs))
done = true
for _, result := range b.Results {
var sentinelKey roachpb.Key
for _, kv := range result.Rows {
// Still processing table.
done = false
if nonNullableColumn != "" {
return fmt.Errorf("column %s contains null values", nonNullableColumn)
}
if sentinelKey == nil || !bytes.HasPrefix(kv.Key, sentinelKey) {
// Sentinel keys have a 0 suffix indicating 0 bytes of
// column ID. Strip off that suffix to determine the
// prefix shared with the other keys for the row.
sentinelKey = sqlbase.StripColumnIDLength(kv.Key)
// Store away key for the next table row as the point from
// which to start from.
curSentinel = sentinelKey
// Delete the entire dropped columns. This used to use SQL
// UPDATE in the past to update the dropped column to
// NULL; but a column in the process of being dropped is
// placed in the table descriptor mutations, and a SQL
// UPDATE of a column in mutations will fail.
for _, columnDesc := range dropped {
// Delete the dropped column.
colKey := keys.MakeColumnKey(sentinelKey, uint32(columnDesc.ID))
if log.V(2) {
log.Infof("Del %s", colKey)
}
writeBatch.Del(colKey)
}
// Add the new columns and backfill the values.
for i, expr := range defaultExprs {
if expr == nil {
continue
}
col := added[i]
colKey := keys.MakeColumnKey(sentinelKey, uint32(col.ID))
d, err := expr.Eval(evalCtx)
if err != nil {
return err
}
marshalled[i], err = sqlbase.MarshalColumnValue(col, d)
if err != nil {
return err
}
if log.V(2) {
log.Infof("Put %s -> %v", colKey, d)
}
// Insert default value into the column. If this row
// was recently added the default value might have
// already been populated, because the
// ColumnDescriptor is in the WRITE_ONLY state.
// Reinserting the default value is not a big deal.
//
// Note: a column in the WRITE_ONLY state cannot be
// populated directly through SQL. A SQL INSERT cannot
// directly reference the column, and the INSERT
// populates the column with the default value.
writeBatch.Put(colKey, &marshalled[i])
}
}
}
}
if err := txn.Run(writeBatch); err != nil {
for _, r := range writeBatch.Results {
if r.PErr != nil {
return convertBackfillError(tableDesc, writeBatch, r.PErr)
}
}
return err
}
return nil
})
return curSentinel.PrefixEnd(), done, err
}
// Add the specified timestamp to the cache as covering the range of
// keys from start to end. If end is nil, the range covers the start
// key only. txnID is nil for no transaction. readTSCache specifies
// whether the command adding this timestamp should update the read
// timestamp; false to update the write timestamp cache.
func (tc *TimestampCache) Add(start, end roachpb.Key, timestamp roachpb.Timestamp, txnID *uuid.UUID, readTSCache bool) {
// This gives us a memory-efficient end key if end is empty.
if len(end) == 0 {
end = start.Next()
start = end[:len(start)]
}
if tc.latest.Less(timestamp) {
tc.latest = timestamp
}
// Only add to the cache if the timestamp is more recent than the
// low water mark.
if tc.lowWater.Less(timestamp) {
cache := tc.wCache
if readTSCache {
cache = tc.rCache
}
addRange := func(r interval.Range) {
value := cacheValue{timestamp: timestamp, txnID: txnID}
key := cache.MakeKey(r.Start, r.End)
entry := makeCacheEntry(key, value)
cache.AddEntry(entry)
}
r := interval.Range{
Start: interval.Comparable(start),
End: interval.Comparable(end),
}
// Check existing, overlapping entries and truncate/split/remove if
// superseded and in the past. If existing entries are in the future,
// subtract from the range/ranges that need to be added to cache.
for _, o := range cache.GetOverlaps(r.Start, r.End) {
cv := o.Value.(*cacheValue)
sCmp := r.Start.Compare(o.Key.Start)
eCmp := r.End.Compare(o.Key.End)
if !timestamp.Less(cv.timestamp) {
// The existing interval has a timestamp less than or equal to the new interval.
// Compare interval ranges to determine how to modify existing interval.
switch {
case sCmp == 0 && eCmp == 0:
// New and old are equal; replace old with new and avoid the need to insert new.
//
// New: ------------
// Old: ------------
//
// New: ------------
*cv = cacheValue{timestamp: timestamp, txnID: txnID}
cache.MoveToEnd(o.Entry)
return
case sCmp <= 0 && eCmp >= 0:
// New contains or is equal to old; delete old.
//
// New: ------------ ------------ ------------
// Old: -------- or ---------- or ----------
//
// Old:
cache.DelEntry(o.Entry)
case sCmp > 0 && eCmp < 0:
// Old contains new; split up old into two.
//
// New: ----
// Old: ------------
//
// Old: ---- ----
oldEnd := o.Key.End
o.Key.End = r.Start
key := cache.MakeKey(r.End, oldEnd)
entry := makeCacheEntry(key, *cv)
cache.AddEntryAfter(entry, o.Entry)
case eCmp >= 0:
// Left partial overlap; truncate old end.
//
// New: -------- --------
// Old: -------- or ------------
//
// Old: ---- ----
o.Key.End = r.Start
case sCmp <= 0:
// Right partial overlap; truncate old start.
//
// New: -------- --------
// Old: -------- or ------------
//
// Old: ---- ----
o.Key.Start = r.End
default:
panic(fmt.Sprintf("no overlap between %v and %v", o.Key.Range, r))
}
} else {
// The existing interval has a timestamp greater than the new interval.
// Compare interval ranges to determine how to modify new interval before
// adding it to the timestamp cache.
switch {
case sCmp >= 0 && eCmp <= 0:
// Old contains or is equal to new; no need to add.
//
// Old: ----------- ----------- ----------- -----------
// New: ----- or ----------- or -------- or --------
//
// New:
return
case sCmp < 0 && eCmp > 0:
// New contains old; split up old into two. We can add the left piece
// immediately because it is guaranteed to be before the rest of the
// overlaps.
//
// Old: ------
// New: ------------
//
// New: --- ---
lr := interval.Range{Start: r.Start, End: o.Key.Start}
addRange(lr)
r.Start = o.Key.End
case eCmp > 0:
// Left partial overlap; truncate new start.
//
// Old: -------- --------
// New: -------- or ------------
//
// New: ---- ----
r.Start = o.Key.End
case sCmp < 0:
// Right partial overlap; truncate new end.
//
// Old: -------- --------
// New: -------- or ------------
//
// New: ---- ----
r.End = o.Key.Start
default:
panic(fmt.Sprintf("no overlap between %v and %v", o.Key.Range, r))
}
}
}
addRange(r)
}
}
