func numComponents(fq *ds.FinalizedQuery) int {
numComponents := len(fq.Orders())
if p, _, _ := fq.IneqFilterLow(); p != "" {
numComponents++
}
if p, _, _ := fq.IneqFilterHigh(); p != "" {
numComponents++
}
for _, v := range fq.EqFilters() {
numComponents += v.Len()
}
return numComponents
}
// adjustQuery applies various mutations to the query to make it suitable for
// merging. In general, this removes limits and offsets the 'distinct' modifier,
// and it ensures that if there are sort orders which won't appear in the
// result data that the query is transformed into a projection query which
// contains all of the data. A non-projection query will never be transformed
// in this way.
func adjustQuery(fq *ds.FinalizedQuery) (*ds.FinalizedQuery, error) {
q := fq.Original()
// The limit and offset must be done in-memory because otherwise we may
// request too few entities from the underlying store if many matching
// entities have been deleted in the buffered transaction.
q = q.Limit(-1)
q = q.Offset(-1)
// distinction must be done in-memory, because otherwise there's no way
// to merge in the effect of the in-flight changes (because there's no way
// to push back to the datastore "yeah, I know you told me that the (1, 2)
// result came from `/Bob,1`, but would you mind pretending that it didn't
// and tell me next the one instead?
q = q.Distinct(false)
// since we need to merge results, we must have all order-related fields
// in each result. The only time we wouldn't have all the data available would
// be for a keys-only or projection query. To fix this, we convert all
// Projection and KeysOnly queries to project on /all/ Orders.
//
// FinalizedQuery already guarantees that all projected fields show up in
// the Orders, but the projected fields could be a subset of the orders.
//
// Additionally on a keys-only query, any orders other than __key__ require
// conversion of this query to a projection query including those orders in
// order to merge the results correctly.
//
// In both cases, the resulting objects returned to the higher layers of the
// stack will only include the information requested by the user; keys-only
// queries will discard all PropertyMap data, and projection queries will
// discard any field data that the user didn't ask for.
orders := fq.Orders()
if len(fq.Project()) > 0 || (fq.KeysOnly() && len(orders) > 1) {
q = q.KeysOnly(false)
for _, o := range orders {
if o.Property == "__key__" {
continue
}
q = q.Project(o.Property)
}
}
return q.Finalize()
}
// GetBinaryBounds gets the binary encoding of the upper and lower bounds of
// the inequality filter on fq, if any is defined. If a bound does not exist,
// it is nil.
//
// NOTE: if fq specifies a descending sort order for the inequality, the bounds
// will be inverted, incremented, and flipped.
func GetBinaryBounds(fq *ds.FinalizedQuery) (lower, upper []byte) {
// Pick up the start/end range from the inequalities, if any.
//
// start and end in the reducedQuery are normalized so that `start >=
// X < end`. Because of that, we need to tweak the inequality filters
// contained in the query if they use the > or <= operators.
if ineqProp := fq.IneqFilterProp(); ineqProp != "" {
_, startOp, startV := fq.IneqFilterLow()
if startOp != "" {
lower = serialize.ToBytes(startV)
if startOp == ">" {
lower = increment(lower)
}
}
_, endOp, endV := fq.IneqFilterHigh()
if endOp != "" {
upper = serialize.ToBytes(endV)
if endOp == "<=" {
upper = increment(upper)
}
}
// The inequality is specified in natural (ascending) order in the query's
// Filter syntax, but the order information may indicate to use a descending
// index column for it. If that's the case, then we must invert, swap and
// increment the inequality endpoints.
//
// Invert so that the desired numbers are represented correctly in the index.
// Swap so that our iterators still go from >= start to < end.
// Increment so that >= and < get correctly bounded (since the iterator is
// still using natrual bytes ordering)
if fq.Orders()[0].Descending {
hi, lo := []byte(nil), []byte(nil)
if len(lower) > 0 {
lo = increment(serialize.Invert(lower))
}
if len(upper) > 0 {
hi = increment(serialize.Invert(upper))
}
upper, lower = lo, hi
}
}
return
}
func (d rdsImpl) fixQuery(fq *ds.FinalizedQuery) (*datastore.Query, error) {
ret := datastore.NewQuery(fq.Kind())
start, end := fq.Bounds()
if start != nil {
ret = ret.Start(start.(datastore.Cursor))
}
if end != nil {
ret = ret.End(end.(datastore.Cursor))
}
for prop, vals := range fq.EqFilters() {
if prop == "__ancestor__" {
p, err := dsF2RProp(d.aeCtx, vals[0])
if err != nil {
return nil, err
}
ret = ret.Ancestor(p.Value.(*datastore.Key))
} else {
filt := prop + "="
for _, v := range vals {
p, err := dsF2RProp(d.aeCtx, v)
if err != nil {
return nil, err
}
ret = ret.Filter(filt, p.Value)
}
}
}
if lnam, lop, lprop := fq.IneqFilterLow(); lnam != "" {
p, err := dsF2RProp(d.aeCtx, lprop)
if err != nil {
return nil, err
}
ret = ret.Filter(lnam+" "+lop, p.Value)
}
if hnam, hop, hprop := fq.IneqFilterHigh(); hnam != "" {
p, err := dsF2RProp(d.aeCtx, hprop)
if err != nil {
return nil, err
}
ret = ret.Filter(hnam+" "+hop, p.Value)
}
if fq.EventuallyConsistent() {
ret = ret.EventualConsistency()
}
if fq.KeysOnly() {
ret = ret.KeysOnly()
}
if lim, ok := fq.Limit(); ok {
ret = ret.Limit(int(lim))
}
if off, ok := fq.Offset(); ok {
ret = ret.Offset(int(off))
}
for _, o := range fq.Orders() {
ret = ret.Order(o.String())
}
ret = ret.Project(fq.Project()...)
if fq.Distinct() {
ret = ret.Distinct()
}
return ret, nil
}
func reduce(fq *ds.FinalizedQuery, aid, ns string, isTxn bool) (*reducedQuery, error) {
if err := fq.Valid(aid, ns); err != nil {
return nil, err
}
if isTxn && fq.Ancestor() == nil {
return nil, fmt.Errorf("queries within a transaction must include an Ancestor filter")
}
if num := numComponents(fq); num > MaxQueryComponents {
return nil, fmt.Errorf(
"gae/memory: query is too large. may not have more than "+
"%d filters + sort orders + ancestor total: had %d",
MaxQueryComponents, num)
}
ret := &reducedQuery{
aid: aid,
ns: ns,
kind: fq.Kind(),
suffixFormat: fq.Orders(),
}
eqFilts := fq.EqFilters()
ret.eqFilters = make(map[string]stringset.Set, len(eqFilts))
for prop, vals := range eqFilts {
sVals := stringset.New(len(vals))
for _, v := range vals {
sVals.Add(string(serialize.ToBytes(v)))
}
ret.eqFilters[prop] = sVals
}
startD, endD := GetBinaryBounds(fq)
// Now we check the start and end cursors.
//
// Cursors are composed of a list of IndexColumns at the beginning, followed
// by the raw bytes to use for the suffix. The cursor is only valid if all of
// its IndexColumns match our proposed suffixFormat, as calculated above.
//
// Cursors are mutually exclusive with the start/end we picked up from the
// inequality. In a well formed query, they indicate a subset of results
// bounded by the inequality. Technically if the start cursor is not >= the
// low bound, or the end cursor is < the high bound, it's an error, but for
// simplicity we just cap to the narrowest intersection of the inequality and
// cursors.
ret.start = startD
ret.end = endD
if start, end := fq.Bounds(); start != nil || end != nil {
if start != nil {
if c, ok := start.(queryCursor); ok {
startCols, startD, err := c.decode()
if err != nil {
return nil, err
}
if !sortOrdersEqual(startCols, ret.suffixFormat) {
return nil, errors.New("gae/memory: start cursor is invalid for this query")
}
if ret.start == nil || bytes.Compare(ret.start, startD) < 0 {
ret.start = startD
}
} else {
return nil, errors.New("gae/memory: bad cursor type")
}
}
if end != nil {
if c, ok := end.(queryCursor); ok {
endCols, endD, err := c.decode()
if err != nil {
return nil, err
}
if !sortOrdersEqual(endCols, ret.suffixFormat) {
return nil, errors.New("gae/memory: end cursor is invalid for this query")
}
if ret.end == nil || bytes.Compare(endD, ret.end) < 0 {
ret.end = endD
}
} else {
return nil, errors.New("gae/memory: bad cursor type")
}
}
}
// Finally, verify that we could even /potentially/ do work. If we have
// overlapping range ends, then we don't have anything to do.
if ret.end != nil && bytes.Compare(ret.start, ret.end) >= 0 {
return nil, ds.ErrNullQuery
}
ret.numCols = len(ret.suffixFormat)
for prop, vals := range ret.eqFilters {
if len(ret.suffixFormat) == 1 && prop == "__ancestor__" {
continue
}
ret.numCols += vals.Len()
}
return ret, nil
}
// runMergedQueries executes a user query `fq` against the parent datastore as
// well as the in-memory datastore, calling `cb` with the merged result set.
//
// It's expected that the caller of this function will apply limit and offset
// if the query contains those restrictions. This may convert the query to
// an expanded projection query with more data than the user asked for. It's the
// caller's responsibility to prune away the extra data.
//
// See also `dsTxnBuf.Run()`.
func runMergedQueries(fq *ds.FinalizedQuery, sizes *sizeTracker,
memDS, parentDS ds.RawInterface, cb func(k *ds.Key, data ds.PropertyMap) error) error {
toRun, err := adjustQuery(fq)
if err != nil {
return err
}
cmpLower, cmpUpper := memory.GetBinaryBounds(fq)
cmpOrder := fq.Orders()
cmpFn := func(i *item) string {
return i.getCmpRow(cmpLower, cmpUpper, cmpOrder)
}
dedup := stringset.Set(nil)
distinct := stringset.Set(nil)
distinctOrder := []ds.IndexColumn(nil)
if len(fq.Project()) > 0 { // the original query was a projection query
if fq.Distinct() {
// it was a distinct projection query, so we need to dedup by distinct
// options.
distinct = stringset.New(0)
proj := fq.Project()
distinctOrder = make([]ds.IndexColumn, len(proj))
for i, p := range proj {
distinctOrder[i].Property = p
}
}
} else {
// the original was a normal or keys-only query, so we need to dedup by keys.
dedup = stringset.New(0)
}
stopChan := make(chan struct{})
parIter := queryToIter(stopChan, toRun, parentDS)
memIter := queryToIter(stopChan, toRun, memDS)
parItemGet := func() (*item, error) {
for {
itm, err := parIter()
if itm == nil || err != nil {
return nil, err
}
encKey := itm.getEncKey()
if sizes.has(encKey) || (dedup != nil && dedup.Has(encKey)) {
continue
}
return itm, nil
}
}
memItemGet := func() (*item, error) {
for {
itm, err := memIter()
if itm == nil || err != nil {
return nil, err
}
if dedup != nil && dedup.Has(itm.getEncKey()) {
continue
}
return itm, nil
}
}
defer func() {
close(stopChan)
parItemGet()
memItemGet()
}()
pitm, err := parItemGet()
if err != nil {
return err
}
mitm, err := memItemGet()
if err != nil {
return err
}
for {
// the err can be set during the loop below. If we come around the bend and
// it's set, then we need to return it. We don't check it immediately
// because it's set after we already have a good result to return to the
// user.
if err != nil {
return err
}
usePitm := pitm != nil
if pitm != nil && mitm != nil {
usePitm = cmpFn(pitm) < cmpFn(mitm)
} else if pitm == nil && mitm == nil {
break
}
toUse := (*item)(nil)
// we check the error at the beginning of the loop.
if usePitm {
toUse = pitm
pitm, err = parItemGet()
} else {
toUse = mitm
mitm, err = memItemGet()
}
if dedup != nil {
if !dedup.Add(toUse.getEncKey()) {
continue
}
}
if distinct != nil {
// NOTE: We know that toUse will not be used after this point for
// comparison purposes, so re-use its cmpRow property for our distinct
// filter here.
toUse.cmpRow = ""
if !distinct.Add(toUse.getCmpRow(nil, nil, distinctOrder)) {
continue
}
}
if err := cb(toUse.key, toUse.data); err != nil {
if err == ds.Stop {
return nil
}
return err
}
}
return nil
}
