// 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.ShallowNext()
start = end[:len(start)]
}
keyR := interval.Range{
Start: interval.Comparable(start),
End: interval.Comparable(end),
}
keys.Add(keyR)
}
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
}
// 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.ShallowNext()
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// superseded and in the past.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)
}
}
// fetch retrieves spans from the kv
func (f *kvFetcher) fetch() *roachpb.Error {
// Retrieve all the spans.
batchSize := int64(kvBatchSize)
if f.firstBatchLimit != 0 && len(f.kvs) == 0 && f.firstBatchLimit < batchSize {
batchSize = f.firstBatchLimit
}
b := &client.Batch{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.ShallowNext()
}
}
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 pErr := f.txn.Run(b); pErr != nil {
return pErr
}
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.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
}