func TestTruncate(t *testing.T) {
defer leaktest.AfterTest(t)()
loc := func(s string) string {
return string(keys.RangeDescriptorKey(roachpb.RKey(s)))
}
locPrefix := func(s string) string {
return string(keys.MakeRangeKeyPrefix(roachpb.RKey(s)))
}
testCases := []struct {
keys [][2]string
expKeys [][2]string
from, to string
desc [2]string // optional, defaults to {from,to}
err string
}{
{
// Keys inside of active range.
keys: [][2]string{{"a", "q"}, {"c"}, {"b, e"}, {"q"}},
expKeys: [][2]string{{"a", "q"}, {"c"}, {"b, e"}, {"q"}},
from: "a", to: "q\x00",
},
{
// Keys outside of active range.
keys: [][2]string{{"a"}, {"a", "b"}, {"q"}, {"q", "z"}},
expKeys: [][2]string{{}, {}, {}, {}},
from: "b", to: "q",
},
{
// Range-local keys inside of active range.
keys: [][2]string{{loc("b")}, {loc("c")}},
expKeys: [][2]string{{loc("b")}, {loc("c")}},
from: "b", to: "e",
},
{
// Range-local key outside of active range.
keys: [][2]string{{loc("a")}},
expKeys: [][2]string{{}},
from: "b", to: "e",
},
{
// Range-local range contained in active range.
keys: [][2]string{{loc("b"), loc("e") + "\x00"}},
expKeys: [][2]string{{loc("b"), loc("e") + "\x00"}},
from: "b", to: "e\x00",
},
{
// Range-local range not contained in active range.
keys: [][2]string{{loc("a"), loc("b")}},
expKeys: [][2]string{{}},
from: "c", to: "e",
},
{
// Range-local range not contained in active range.
keys: [][2]string{{loc("a"), locPrefix("b")}, {loc("e"), loc("f")}},
expKeys: [][2]string{{}, {}},
from: "b", to: "e",
},
{
// Range-local range partially contained in active range.
keys: [][2]string{{loc("a"), loc("b")}},
expKeys: [][2]string{{loc("a"), locPrefix("b")}},
from: "a", to: "b",
},
{
// Range-local range partially contained in active range.
keys: [][2]string{{loc("a"), loc("b")}},
expKeys: [][2]string{{locPrefix("b"), loc("b")}},
from: "b", to: "e",
},
{
// Range-local range contained in active range.
keys: [][2]string{{locPrefix("b"), loc("b")}},
expKeys: [][2]string{{locPrefix("b"), loc("b")}},
from: "b", to: "c",
},
{
// Mixed range-local vs global key range.
keys: [][2]string{{loc("c"), "d\x00"}},
from: "b", to: "e",
err: "local key mixed with global key",
},
{
// Key range touching and intersecting active range.
keys: [][2]string{{"a", "b"}, {"a", "c"}, {"p", "q"}, {"p", "r"}, {"a", "z"}},
expKeys: [][2]string{{}, {"b", "c"}, {"p", "q"}, {"p", "q"}, {"b", "q"}},
from: "b", to: "q",
},
// Active key range is intersection of descriptor and [from,to).
{
keys: [][2]string{{"c", "q"}},
expKeys: [][2]string{{"d", "p"}},
from: "a", to: "z",
desc: [2]string{"d", "p"},
},
{
keys: [][2]string{{"c", "q"}},
expKeys: [][2]string{{"d", "p"}},
from: "d", to: "p",
desc: [2]string{"a", "z"},
},
}
for i, test := range testCases {
goldenOriginal := roachpb.BatchRequest{}
for _, ks := range test.keys {
if len(ks[1]) > 0 {
u := uuid.MakeV4()
goldenOriginal.Add(&roachpb.ResolveIntentRangeRequest{
Span: roachpb.Span{Key: roachpb.Key(ks[0]), EndKey: roachpb.Key(ks[1])},
IntentTxn: enginepb.TxnMeta{ID: &u},
})
} else {
goldenOriginal.Add(&roachpb.GetRequest{
Span: roachpb.Span{Key: roachpb.Key(ks[0])},
})
}
}
original := roachpb.BatchRequest{Requests: make([]roachpb.RequestUnion, len(goldenOriginal.Requests))}
for i, request := range goldenOriginal.Requests {
original.Requests[i].SetValue(request.GetInner().ShallowCopy())
}
desc := &roachpb.RangeDescriptor{
StartKey: roachpb.RKey(test.desc[0]), EndKey: roachpb.RKey(test.desc[1]),
}
if len(desc.StartKey) == 0 {
desc.StartKey = roachpb.RKey(test.from)
}
if len(desc.EndKey) == 0 {
desc.EndKey = roachpb.RKey(test.to)
}
rs := roachpb.RSpan{Key: roachpb.RKey(test.from), EndKey: roachpb.RKey(test.to)}
rs, err := rs.Intersect(desc)
if err != nil {
t.Errorf("%d: intersection failure: %v", i, err)
continue
}
ba, num, err := truncate(original, rs)
if err != nil || test.err != "" {
if !testutils.IsError(err, test.err) {
t.Errorf("%d: %v (expected: %q)", i, err, test.err)
}
continue
}
var reqs int
for j, arg := range ba.Requests {
req := arg.GetInner()
if _, ok := req.(*roachpb.NoopRequest); ok {
continue
}
if h := req.Header(); !bytes.Equal(h.Key, roachpb.Key(test.expKeys[j][0])) || !bytes.Equal(h.EndKey, roachpb.Key(test.expKeys[j][1])) {
t.Errorf("%d.%d: range mismatch: actual [%q,%q), wanted [%q,%q)", i, j,
h.Key, h.EndKey, test.expKeys[j][0], test.expKeys[j][1])
} else if _, ok := req.(*roachpb.NoopRequest); ok != (len(h.Key) == 0) {
t.Errorf("%d.%d: expected NoopRequest, got %T", i, j, req)
} else if len(h.Key) != 0 {
reqs++
}
}
if reqs != num {
t.Errorf("%d: counted %d requests, but truncation indicated %d", i, reqs, num)
}
if !reflect.DeepEqual(original, goldenOriginal) {
t.Errorf("%d: truncation mutated original:\nexpected: %s\nactual: %s",
i, goldenOriginal, original)
}
}
}
// truncate restricts all contained requests to the given key range
// and returns a new BatchRequest.
// All requests contained in that batch are "truncated" to the given
// span, 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.
func truncate(ba roachpb.BatchRequest, rs roachpb.RSpan) (roachpb.BatchRequest, int, error) {
truncateOne := func(args roachpb.Request) (bool, roachpb.Span, error) {
if _, ok := args.(*roachpb.NoopRequest); ok {
return true, emptySpan, nil
}
header := args.Header()
if !roachpb.IsRange(args) {
// This is a point request.
if len(header.EndKey) > 0 {
return false, emptySpan, errors.Errorf("%T is not a range command, but EndKey is set", args)
}
keyAddr, err := keys.Addr(header.Key)
if err != nil {
return false, emptySpan, err
}
if !rs.ContainsKey(keyAddr) {
return false, emptySpan, nil
}
return true, header, nil
}
// We're dealing with a range-spanning request.
local := false
keyAddr, err := keys.Addr(header.Key)
if err != nil {
return false, emptySpan, err
}
endKeyAddr, err := keys.Addr(header.EndKey)
if err != nil {
return false, emptySpan, err
}
if l, r := !keyAddr.Equal(header.Key), !endKeyAddr.Equal(header.EndKey); l || r {
if !l || !r {
return false, emptySpan, errors.Errorf("local key mixed with global key in range")
}
local = true
}
if keyAddr.Less(rs.Key) {
// rs.Key can't be local because it contains range split points, which
// are never local.
if !local {
header.Key = rs.Key.AsRawKey()
} else {
// The local start key should be truncated to the boundary of local keys which
// address to rs.Key.
header.Key = keys.MakeRangeKeyPrefix(rs.Key)
}
}
if !endKeyAddr.Less(rs.EndKey) {
// rs.EndKey can't be local because it contains range split points, which
// are never local.
if !local {
header.EndKey = rs.EndKey.AsRawKey()
} else {
// The local end key should be truncated to the boundary of local keys which
// address to rs.EndKey.
header.EndKey = keys.MakeRangeKeyPrefix(rs.EndKey)
}
}
// Check whether the truncation has left any keys in the range. If not,
// we need to cut it out of the request.
if header.Key.Compare(header.EndKey) >= 0 {
return false, emptySpan, nil
}
return true, header, nil
}
var numNoop int
truncBA := ba
truncBA.Requests = make([]roachpb.RequestUnion, len(ba.Requests))
for pos, arg := range ba.Requests {
hasRequest, newHeader, err := truncateOne(arg.GetInner())
if !hasRequest {
// We omit this one, i.e. replace it with a Noop.
numNoop++
union := roachpb.RequestUnion{}
union.MustSetInner(&noopRequest)
truncBA.Requests[pos] = union
} else {
// Keep the old one. If we must adjust the header, must copy.
if inner := ba.Requests[pos].GetInner(); newHeader.Equal(inner.Header()) {
truncBA.Requests[pos] = ba.Requests[pos]
} else {
shallowCopy := inner.ShallowCopy()
shallowCopy.SetHeader(newHeader)
union := &truncBA.Requests[pos] // avoid operating on copy
union.MustSetInner(shallowCopy)
}
}
if err != nil {
return roachpb.BatchRequest{}, 0, err
}
}
return truncBA, len(ba.Requests) - numNoop, nil
}
// sendPartialBatch sends the supplied batch to the range specified by
// desc. The batch request is first truncated so that it contains only
// requests which intersect the range descriptor and keys for each
// request are limited to the range's key span. The send occurs in a
// retry loop to handle send failures. On failure to send to any
// replicas, we backoff and retry by refetching the range
// descriptor. If the underlying range seems to have split, we
// recursively invoke divideAndSendBatchToRanges to re-enumerate the
// ranges in the span and resend to each.
func (ds *DistSender) sendPartialBatch(
ctx context.Context,
ba roachpb.BatchRequest,
rs roachpb.RSpan,
desc *roachpb.RangeDescriptor,
evictToken *EvictionToken,
isFirst bool,
) response {
var reply *roachpb.BatchResponse
var pErr *roachpb.Error
isReverse := ba.IsReverse()
// Truncate the request to range descriptor.
intersected, err := rs.Intersect(desc)
if err != nil {
return response{pErr: roachpb.NewError(err)}
}
truncBA, numActive, err := truncate(ba, intersected)
if numActive == 0 && err == nil {
// This shouldn't happen in the wild, but some tests exercise it.
return response{
pErr: roachpb.NewErrorf("truncation resulted in empty batch on %s: %s", intersected, ba),
}
}
if err != nil {
return response{pErr: roachpb.NewError(err)}
}
// Start a retry loop for sending the batch to the range.
for r := retry.StartWithCtx(ctx, ds.rpcRetryOptions); r.Next(); {
// If we've cleared the descriptor on a send failure, re-lookup.
if desc == nil {
var descKey roachpb.RKey
if isReverse {
descKey = intersected.EndKey
} else {
descKey = intersected.Key
}
desc, evictToken, err = ds.getDescriptor(ctx, descKey, nil, isReverse)
if err != nil {
log.ErrEventf(ctx, "range descriptor re-lookup failed: %s", err)
continue
}
}
reply, pErr = ds.sendSingleRange(ctx, truncBA, desc)
// If sending succeeded, return immediately.
if pErr == nil {
return response{reply: reply}
}
log.ErrEventf(ctx, "reply error %s: %s", ba, pErr)
// Error handling: If the error indicates that our range
// descriptor is out of date, evict it from the cache and try
// again. Errors that apply only to a single replica were
// handled in send().
//
// TODO(bdarnell): Don't retry endlessly. If we fail twice in a
// row and the range descriptor hasn't changed, return the error
// to our caller.
switch tErr := pErr.GetDetail().(type) {
case *roachpb.SendError:
// We've tried all the replicas without success. Either
// they're all down, or we're using an out-of-date range
// descriptor. Invalidate the cache and try again with the new
// metadata.
log.Event(ctx, "evicting range descriptor on send error and backoff for re-lookup")
if err := evictToken.Evict(ctx); err != nil {
return response{pErr: roachpb.NewError(err)}
}
// Clear the descriptor to reload on the next attempt.
desc = nil
continue
case *roachpb.RangeKeyMismatchError:
// Range descriptor might be out of date - evict it. This is
// likely the result of a range split. If we have new range
// descriptors, insert them instead as long as they are different
// from the last descriptor to avoid endless loops.
var replacements []roachpb.RangeDescriptor
different := func(rd *roachpb.RangeDescriptor) bool {
return !desc.RSpan().Equal(rd.RSpan())
}
if tErr.MismatchedRange != nil && different(tErr.MismatchedRange) {
replacements = append(replacements, *tErr.MismatchedRange)
}
if tErr.SuggestedRange != nil && different(tErr.SuggestedRange) {
if includesFrontOfCurSpan(isReverse, tErr.SuggestedRange, rs) {
replacements = append(replacements, *tErr.SuggestedRange)
}
}
// Same as Evict() if replacements is empty.
if err := evictToken.EvictAndReplace(ctx, replacements...); err != nil {
return response{pErr: roachpb.NewError(err)}
}
// On addressing errors (likely a split), we need to re-invoke
// the range descriptor lookup machinery, so we recurse by
// sending batch to just the partial span this descriptor was
// supposed to cover.
log.VEventf(ctx, 1, "likely split; resending batch to span: %s", tErr)
reply, pErr = ds.divideAndSendBatchToRanges(ctx, ba, intersected, isFirst)
return response{reply: reply, pErr: pErr}
}
break
}
// Propagate error if either the retry closer or context done
// channels were closed.
if pErr == nil {
if pErr = ds.deduceRetryEarlyExitError(ctx); pErr == nil {
log.Fatal(ctx, "exited retry loop without an error")
}
}
return response{pErr: pErr}
}