func pickQueryStrategy(fq *ds.FinalizedQuery, rq *reducedQuery, cb ds.RawRunCB, head *memStore) queryStrategy {
if fq.KeysOnly() {
return &keysOnlyStrategy{cb, stringset.New(0)}
}
if len(fq.Project()) > 0 {
return newProjectionStrategy(fq, rq, cb)
}
return newNormalStrategy(rq.aid, rq.ns, cb, head)
}
func countQuery(fq *ds.FinalizedQuery, aid, ns string, isTxn bool, idx, head *memStore) (ret int64, err error) {
if len(fq.Project()) == 0 && !fq.KeysOnly() {
fq, err = fq.Original().KeysOnly(true).Finalize()
if err != nil {
return
}
}
err = executeQuery(fq, aid, ns, isTxn, idx, head, func(_ *ds.Key, _ ds.PropertyMap, _ ds.CursorCB) error {
ret++
return nil
})
return
}
func (d *dsTxnBuf) Count(fq *ds.FinalizedQuery) (count int64, err error) {
// Unfortunately there's no fast-path here. We literally have to run the
// query and count. Fortunately we can optimize to count keys if it's not
// a projection query. This will save on bandwidth a bit.
if len(fq.Project()) == 0 && !fq.KeysOnly() {
fq, err = fq.Original().KeysOnly(true).Finalize()
if err != nil {
return
}
}
err = d.Run(fq, func(_ *ds.Key, _ ds.PropertyMap, _ ds.CursorCB) error {
count++
return nil
})
return
}
// 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()
}
func (d *dsTxnBuf) Run(fq *ds.FinalizedQuery, cb ds.RawRunCB) error {
if start, end := fq.Bounds(); start != nil || end != nil {
return errors.New("txnBuf filter does not support query cursors")
}
limit, limitSet := fq.Limit()
offset, _ := fq.Offset()
keysOnly := fq.KeysOnly()
project := fq.Project()
bufDS, parentDS, sizes := func() (ds.RawInterface, ds.RawInterface, *sizeTracker) {
if !d.haveLock {
d.state.Lock()
defer d.state.Unlock()
}
return d.state.bufDS, d.state.parentDS, d.state.entState.dup()
}()
return runMergedQueries(fq, sizes, bufDS, parentDS, func(key *ds.Key, data ds.PropertyMap) error {
if offset > 0 {
offset--
return nil
}
if limitSet {
if limit == 0 {
return ds.Stop
}
limit--
}
if keysOnly {
data = nil
} else if len(project) > 0 {
newData := make(ds.PropertyMap, len(project))
for _, p := range project {
newData[p] = data[p]
}
data = newData
}
return cb(key, data, nil)
})
}
func newProjectionStrategy(fq *ds.FinalizedQuery, rq *reducedQuery, cb ds.RawRunCB) queryStrategy {
proj := fq.Project()
projectionLookups := make([]projectionLookup, len(proj))
for i, prop := range proj {
projectionLookups[i].propertyName = prop
lookupErr := fmt.Errorf("planning a strategy for an unfulfillable query?")
for j, col := range rq.suffixFormat {
if col.Property == prop {
projectionLookups[i].suffixIndex = j
lookupErr = nil
break
}
}
impossible(lookupErr)
}
ret := &projectionStrategy{cb: cb, project: projectionLookups}
if fq.Distinct() {
ret.distinct = stringset.New(0)
}
return ret
}
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
}
// 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
}