// 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)
}
}
// divideAndSendBatchToRanges sends the supplied batch to all of the
// ranges which comprise the span specified by rs. The batch request
// is trimmed against each range which is part of the span and sent
// either serially or in parallel, if possible. isFirst indicates
// whether this is the first time this method has been called on the
// batch. It's specified false where this method is invoked recursively.
func (ds *DistSender) divideAndSendBatchToRanges(
ctx context.Context, ba roachpb.BatchRequest, rs roachpb.RSpan, isFirst bool,
) (br *roachpb.BatchResponse, pErr *roachpb.Error) {
// This function builds a channel of responses for each range
// implicated in the span (rs) and combines them into a single
// BatchResponse when finished.
var responseChs []chan response
defer func() {
for _, responseCh := range responseChs {
resp := <-responseCh
if resp.pErr != nil {
if pErr == nil {
pErr = resp.pErr
}
continue
}
if br == nil {
// First response from a Range.
br = resp.reply
} else {
// This was the second or later call in a cross-Range request.
// Combine the new response with the existing one.
if err := br.Combine(resp.reply); err != nil {
pErr = roachpb.NewError(err)
return
}
br.Txn.Update(resp.reply.Txn)
}
}
// If we experienced an error, don't neglect to update the error's
// attached transaction with any responses which were received.
if pErr != nil {
if br != nil {
pErr.UpdateTxn(br.Txn)
}
}
}()
// Get initial seek key depending on direction of iteration.
var seekKey roachpb.RKey
isReverse := ba.IsReverse()
if isReverse {
seekKey = rs.EndKey
} else {
seekKey = rs.Key
}
// Send the request to one range per iteration.
ri := NewRangeIterator(ds, isReverse)
for ri.Seek(ctx, seekKey); ri.Valid(); ri.Seek(ctx, seekKey) {
// Increase the sequence counter only once before sending RPCs to
// the ranges involved in this chunk of the batch (as opposed to
// for each RPC individually). On RPC errors, there's no guarantee
// that the request hasn't made its way to the target regardless
// of the error; we'd like the second execution to be caught by
// the sequence cache if that happens. There is a small chance
// that we address a range twice in this chunk (stale/suboptimal
// descriptors due to splits/merges) which leads to a transaction
// retry.
//
// TODO(tschottdorf): it's possible that if we don't evict from
// the cache we could be in for a busy loop.
ba.SetNewRequest()
responseCh := make(chan response, 1)
responseChs = append(responseChs, responseCh)
if isFirst && ri.NeedAnother(rs) {
// TODO(tschottdorf): we should have a mechanism for discovering
// range merges (descriptor staleness will mostly go unnoticed),
// or we'll be turning single-range queries into multi-range
// queries for no good reason.
//
// If there's no transaction and op spans ranges, possibly
// re-run as part of a transaction for consistency. The
// case where we don't need to re-run is if the read
// consistency is not required.
if ba.Txn == nil && ba.IsPossibleTransaction() && ba.ReadConsistency != roachpb.INCONSISTENT {
responseCh <- response{pErr: roachpb.NewError(&roachpb.OpRequiresTxnError{})}
return
}
// If the request is more than but ends with EndTransaction, we
// want the caller to come again with the EndTransaction in an
// extra call.
if l := len(ba.Requests) - 1; l > 0 && ba.Requests[l].GetInner().Method() == roachpb.EndTransaction {
responseCh <- response{pErr: errNo1PCTxn}
return
}
}
// Determine next seek key, taking a potentially sparse batch into
// consideration.
var err error
nextRS := rs
if isReverse {
// In next iteration, query previous range.
// We use the StartKey of the current descriptor as opposed to the
// EndKey of the previous one since that doesn't have bugs when
// stale descriptors come into play.
seekKey, err = prev(ba, ri.Desc().StartKey)
nextRS.EndKey = seekKey
} else {
// In next iteration, query next range.
// It's important that we use the EndKey of the current descriptor
// as opposed to the StartKey of the next one: if the former is stale,
// it's possible that the next range has since merged the subsequent
// one, and unless both descriptors are stale, the next descriptor's
// StartKey would move us to the beginning of the current range,
// resulting in a duplicate scan.
seekKey, err = next(ba, ri.Desc().EndKey)
nextRS.Key = seekKey
}
if err != nil {
responseCh <- response{pErr: roachpb.NewError(err)}
return
}
// Send the next partial batch to the first range in the "rs" span.
// If we're not handling a request which limits responses and we
// can reserve one of the limited goroutines available for parallel
// batch RPCs, send asynchronously.
if ba.MaxSpanRequestKeys == 0 && ri.NeedAnother(rs) && ds.rpcContext != nil &&
ds.sendPartialBatchAsync(ctx, ba, rs, ri.Desc(), ri.Token(), isFirst, responseCh) {
// Note that we pass the batch request by value to the parallel
// goroutine to avoid using the cloned txn.
// Clone the txn to preserve the current txn sequence for the async call.
if ba.Txn != nil {
txnClone := ba.Txn.Clone()
ba.Txn = &txnClone
}
} else {
// Send synchronously if there is no parallel capacity left, there's a
// max results limit, or this is the final request in the span.
resp := ds.sendPartialBatch(ctx, ba, rs, ri.Desc(), ri.Token(), isFirst)
responseCh <- resp
if resp.pErr != nil {
return
}
ba.UpdateTxn(resp.reply.Txn)
// Check whether we've received enough responses to exit query loop.
if ba.MaxSpanRequestKeys > 0 {
var numResults int64
for _, r := range resp.reply.Responses {
numResults += r.GetInner().Header().NumKeys
}
if numResults > ba.MaxSpanRequestKeys {
panic(fmt.Sprintf("received %d results, limit was %d", numResults, ba.MaxSpanRequestKeys))
}
ba.MaxSpanRequestKeys -= numResults
// Exiting; fill in missing responses.
if ba.MaxSpanRequestKeys == 0 {
fillSkippedResponses(ba, resp.reply, seekKey)
return
}
}
}
// Check for completion.
if !ri.NeedAnother(rs) {
return
}
isFirst = false // next range will not be first!
rs = nextRS
}
// We've exited early. Return the range iterator error.
responseCh := make(chan response, 1)
responseCh <- response{pErr: ri.Error()}
responseChs = append(responseChs, responseCh)
return
}
// 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}
}
// initAndVerifyBatch initializes timestamp-related information and
// verifies batch constraints before splitting.
func (ds *DistSender) initAndVerifyBatch(
ctx context.Context, ba *roachpb.BatchRequest,
) *roachpb.Error {
// In the event that timestamp isn't set and read consistency isn't
// required, set the timestamp using the local clock.
if ba.ReadConsistency == roachpb.INCONSISTENT && ba.Timestamp.Equal(hlc.ZeroTimestamp) {
ba.Timestamp = ds.clock.Now()
}
if ba.Txn != nil {
// Make a copy here since the code below modifies it in different places.
// TODO(tschottdorf): be smarter about this - no need to do it for
// requests that don't get split.
txnClone := ba.Txn.Clone()
ba.Txn = &txnClone
if len(ba.Txn.ObservedTimestamps) == 0 {
// Ensure the local NodeID is marked as free from clock offset;
// the transaction's timestamp was taken off the local clock.
if nDesc := ds.getNodeDescriptor(); nDesc != nil {
// TODO(tschottdorf): future refactoring should move this to txn
// creation in TxnCoordSender, which is currently unaware of the
// NodeID (and wraps *DistSender through client.Sender since it
// also needs test compatibility with *LocalSender).
//
// Taking care below to not modify any memory referenced from
// our BatchRequest which may be shared with others.
//
// We already have a clone of our txn (see above), so we can
// modify it freely.
//
// Zero the existing data. That makes sure that if we had
// something of size zero but with capacity, we don't re-use the
// existing space (which others may also use). This is just to
// satisfy paranoia/OCD and not expected to matter in practice.
ba.Txn.ResetObservedTimestamps()
// OrigTimestamp is the HLC timestamp at which the Txn started, so
// this effectively means no more uncertainty on this node.
ba.Txn.UpdateObservedTimestamp(nDesc.NodeID, ba.Txn.OrigTimestamp)
}
}
}
if len(ba.Requests) < 1 {
return roachpb.NewErrorf("empty batch")
}
if ba.MaxSpanRequestKeys != 0 {
// Verify that the batch contains only specific range requests or the
// Begin/EndTransactionRequest. Verify that a batch with a ReverseScan
// only contains ReverseScan range requests.
isReverse := ba.IsReverse()
for _, req := range ba.Requests {
inner := req.GetInner()
switch inner.(type) {
case *roachpb.ScanRequest, *roachpb.DeleteRangeRequest:
// Accepted range requests. All other range requests are still
// not supported.
// TODO(vivek): don't enumerate all range requests.
if isReverse {
return roachpb.NewErrorf("batch with limit contains both forward and reverse scans")
}
case *roachpb.BeginTransactionRequest, *roachpb.EndTransactionRequest, *roachpb.ReverseScanRequest:
continue
default:
return roachpb.NewErrorf("batch with limit contains %T request", inner)
}
}
}
return nil
}