// 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 {
inner := union.GetInner()
if _, ok := inner.(*roachpb.NoopRequest); ok {
continue
}
h := inner.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
}
// 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 {
inner := union.GetInner()
if _, ok := inner.(*roachpb.NoopRequest); ok {
continue
}
h := inner.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
}
// 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)
containsFn := (*roachpb.RangeDescriptor).ContainsKey
if inclusive {
containsFn = (*roachpb.RangeDescriptor).ContainsExclusiveEndKey
}
// Return nil if the key does not belong to the range.
if !containsFn(rd, key) {
return nil, nil, nil
}
return metaEndKey, rd, 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(
ctx context.Context, key roachpb.RKey, desc *roachpb.RangeDescriptor, 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,
Reverse: useReverseScan,
})
replicas := newReplicaSlice(ds.gossip, desc)
replicas.Shuffle()
br, err := ds.sendRPC(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
}
func (m *modelTimeSeriesDataStore) ContainsTimeSeries(start, end roachpb.RKey) bool {
if !start.Less(end) {
m.t.Fatalf("ContainsTimeSeries passed start key %v which is not less than end key %v", start, end)
}
m.Lock()
defer m.Unlock()
m.containsCalled++
return true
}
// 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.DecodeUvarintAscending(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) {
ctx := ls.AnnotateCtx(context.TODO())
log.Warningf(ctx, "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
}
// findTimeSeries searches the supplied engine over the supplied key range,
// identifying time series which have stored data in the range, along with the
// resolutions at which time series data is stored. A unique name/resolution
// pair will only be identified once, even if the range contains keys for that
// name/resolution pair at multiple timestamps or from multiple sources.
//
// An engine snapshot is used, rather than a client, because this function is
// intended to be called by a storage queue which can inspect the local data for
// a single range without the need for expensive network calls.
func findTimeSeries(
snapshot engine.Reader, startKey, endKey roachpb.RKey, now hlc.Timestamp,
) ([]timeSeriesResolutionInfo, error) {
var results []timeSeriesResolutionInfo
iter := snapshot.NewIterator(false)
defer iter.Close()
// Set start boundary for the search, which is the lesser of the range start
// key and the beginning of time series data.
start := engine.MakeMVCCMetadataKey(startKey.AsRawKey())
next := engine.MakeMVCCMetadataKey(keys.TimeseriesPrefix)
if next.Less(start) {
next = start
}
// Set end boundary for the search, which is the lesser of the range end key
// and the end of time series data.
end := engine.MakeMVCCMetadataKey(endKey.AsRawKey())
lastTS := engine.MakeMVCCMetadataKey(keys.TimeseriesPrefix.PrefixEnd())
if lastTS.Less(end) {
end = lastTS
}
thresholds := computeThresholds(now.WallTime)
for iter.Seek(next); iter.Valid() && iter.Less(end); iter.Seek(next) {
foundKey := iter.Key().Key
// Extract the name and resolution from the discovered key.
name, _, res, tsNanos, err := DecodeDataKey(foundKey)
if err != nil {
return nil, err
}
// Skip this time series if there's nothing to prune. We check the
// oldest (first) time series record's timestamp against the
// pruning threshold.
if threshold, ok := thresholds[res]; !ok || threshold > tsNanos {
results = append(results, timeSeriesResolutionInfo{
Name: name,
Resolution: res,
})
}
// Set 'next' is initialized to the next possible time series key
// which could belong to a previously undiscovered time series.
next = engine.MakeMVCCMetadataKey(makeDataKeySeriesPrefix(name, res).PrefixEnd())
}
return results, nil
}
// findTimeSeries searches the supplied engine over the supplied key range,
// identifying time series which have stored data in the range, along with the
// resolutions at which time series data is stored. A unique name/resolution
// pair will only be identified once, even if the range contains keys for that
// name/resolution pair at multiple timestamps or from multiple sources.
//
// An engine snapshot is used, rather than a client, because this function is
// intended to be called by a storage queue which can inspect the local data for
// a single range without the need for expensive network calls.
func findTimeSeries(
snapshot engine.Reader, startKey, endKey roachpb.RKey,
) ([]timeSeriesResolutionInfo, error) {
var results []timeSeriesResolutionInfo
iter := snapshot.NewIterator(false)
defer iter.Close()
// Set start boundary for the search, which is the lesser of the range start
// key and the beginning of time series data.
start := engine.MakeMVCCMetadataKey(startKey.AsRawKey())
next := engine.MakeMVCCMetadataKey(keys.TimeseriesPrefix)
if next.Less(start) {
next = start
}
// Set end boundary for the search, which is the lesser of the range end key
// and the end of time series data.
end := engine.MakeMVCCMetadataKey(endKey.AsRawKey())
lastTS := engine.MakeMVCCMetadataKey(keys.TimeseriesPrefix.PrefixEnd())
if lastTS.Less(end) {
end = lastTS
}
for iter.Seek(next); iter.Valid() && iter.Less(end); iter.Seek(next) {
foundKey := iter.Key().Key
// Extract the name and resolution from the discovered key.
name, _, res, _, err := DecodeDataKey(foundKey)
if err != nil {
return nil, err
}
results = append(results, timeSeriesResolutionInfo{
Name: name,
Resolution: res,
})
// Set 'next' is initialized to the next possible time series key
// which could belong to a previously undiscovered time series.
next = engine.MakeMVCCMetadataKey(makeDataKeySeriesPrefix(name, res).PrefixEnd())
}
return results, 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
}
// 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
}
// 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
}
func (m *modelTimeSeriesDataStore) PruneTimeSeries(
ctx context.Context,
snapshot engine.Reader,
start, end roachpb.RKey,
db *client.DB,
now hlc.Timestamp,
) error {
if snapshot == nil {
m.t.Fatal("PruneTimeSeries was passed a nil snapshot")
}
if db == nil {
m.t.Fatal("PruneTimeSeries was passed a nil client.DB")
}
if !start.Less(end) {
m.t.Fatalf("PruneTimeSeries passed start key %v which is not less than end key %v", start, end)
}
m.Lock()
defer m.Unlock()
m.pruneCalled++
m.pruneSeenStartKeys[start.String()] = struct{}{}
m.pruneSeenEndKeys[end.String()] = struct{}{}
return nil
}
// fillSkippedResponses after meeting the batch key max limit for range
// requests.
func fillSkippedResponses(ba roachpb.BatchRequest, br *roachpb.BatchResponse, nextKey roachpb.RKey) {
// Some requests might have NoopResponses; we must replace them with empty
// responses of the proper type.
for i, req := range ba.Requests {
if _, ok := br.Responses[i].GetInner().(*roachpb.NoopResponse); !ok {
continue
}
var reply roachpb.Response
switch t := req.GetInner().(type) {
case *roachpb.ScanRequest:
reply = &roachpb.ScanResponse{}
case *roachpb.ReverseScanRequest:
reply = &roachpb.ReverseScanResponse{}
case *roachpb.DeleteRangeRequest:
reply = &roachpb.DeleteRangeResponse{}
case *roachpb.BeginTransactionRequest, *roachpb.EndTransactionRequest:
continue
default:
panic(fmt.Sprintf("bad type %T", t))
}
union := roachpb.ResponseUnion{}
union.MustSetInner(reply)
br.Responses[i] = union
}
// Set the ResumeSpan for future batch requests.
isReverse := ba.IsReverse()
for i, resp := range br.Responses {
req := ba.Requests[i].GetInner()
if !roachpb.IsRange(req) {
continue
}
hdr := resp.GetInner().Header()
origSpan := req.Header()
if isReverse {
if hdr.ResumeSpan != nil {
// The ResumeSpan.Key might be set to the StartKey of a range;
// correctly set it to the Key of the original request span.
hdr.ResumeSpan.Key = origSpan.Key
} else if roachpb.RKey(origSpan.Key).Less(nextKey) {
// Some keys have yet to be processed.
hdr.ResumeSpan = &origSpan
if nextKey.Less(roachpb.RKey(origSpan.EndKey)) {
// The original span has been partially processed.
hdr.ResumeSpan.EndKey = nextKey.AsRawKey()
}
}
} else {
if hdr.ResumeSpan != nil {
// The ResumeSpan.EndKey might be set to the EndKey of a
// range; correctly set it to the EndKey of the original
// request span.
hdr.ResumeSpan.EndKey = origSpan.EndKey
} else if nextKey.Less(roachpb.RKey(origSpan.EndKey)) {
// Some keys have yet to be processed.
hdr.ResumeSpan = &origSpan
if roachpb.RKey(origSpan.Key).Less(nextKey) {
// The original span has been partially processed.
hdr.ResumeSpan.Key = nextKey.AsRawKey()
}
}
}
br.Responses[i].GetInner().SetHeader(hdr)
}
}
// ContainsTimeSeries returns true if the given key range overlaps the
// range of possible time series keys.
func (tsdb *DB) ContainsTimeSeries(start, end roachpb.RKey) bool {
return !lastTSRKey.Less(start) && !end.Less(firstTSRKey)
}
// 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 {
tableStart := roachpb.RKey(keys.SystemConfigTableDataMax)
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.MaxSystemConfigDescID {
// 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.MaxSystemConfigDescID + 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.MakeRowSentinelKey(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(context.TODO(), "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(context.TODO(), "unable to determine largest object ID from system config: %s", err)
return nil
}
appendSplitKeys(startID, endID)
return splitKeys
}