ceremonyclient/pebble/metamorphic/key_manager.go

581 lines
19 KiB
Go
Raw Permalink Normal View History

2024-01-03 07:31:42 +00:00
package metamorphic
import (
"cmp"
"fmt"
"slices"
"github.com/cockroachdb/pebble"
"github.com/cockroachdb/pebble/internal/base"
"github.com/cockroachdb/pebble/internal/testkeys"
"github.com/stretchr/testify/require"
)
// objKey is a tuple of (objID, key). This struct is used primarily as a map
// key for keyManager. Only writer objTags can occur here, i.e., dbTag and
// batchTag, since this is used for tracking the keys in a writer.
type objKey struct {
id objID
key []byte
}
// makeObjKey returns a new objKey given and id and key.
func makeObjKey(id objID, key []byte) objKey {
if id.tag() != dbTag && id.tag() != batchTag {
panic("unexpected non-writer tag")
}
return objKey{id, key}
}
// String implements fmt.Stringer, returning a stable string representation of
// the objKey. This string is used as map key.
func (o objKey) String() string {
return fmt.Sprintf("%s:%s", o.id, o.key)
}
type keyUpdate struct {
deleted bool
// metaTimestamp at which the write or delete op occurred.
metaTimestamp int
}
// keyMeta is metadata associated with an (objID, key) pair, where objID is
// a writer containing the key.
type keyMeta struct {
objKey
// The number of Sets of the key in this writer.
sets int
// The number of Merges of the key in this writer.
merges int
// singleDel can be true only if sets <= 1 && merges == 0 and the
// SingleDelete was added to this writer after the set.
singleDel bool
// The number of Deletes of the key in this writer.
dels int
// del can be true only if a Delete was added to this writer after the
// Sets and Merges counted above.
del bool
// updateOps should always be ordered by non-decreasing metaTimestamp.
// updateOps will not be updated if the key is range deleted. Therefore, it
// is a best effort sequence of updates to the key. updateOps is used to
// determine if an iterator created on the DB can read a certain key.
updateOps []keyUpdate
}
func (m *keyMeta) clear() {
m.sets = 0
m.merges = 0
m.singleDel = false
m.del = false
m.dels = 0
m.updateOps = nil
}
// mergeInto merges this metadata this into the metadata for other.
func (m *keyMeta) mergeInto(keyManager *keyManager, other *keyMeta) {
if other.del && !m.del {
// m's Sets and Merges are later.
if m.sets > 0 || m.merges > 0 {
other.del = false
}
} else {
other.del = m.del
}
// Sets, merges, dels are additive.
other.sets += m.sets
other.merges += m.merges
other.dels += m.dels
// Single deletes are preserved. This is valid since we are also
// maintaining a global invariant that SingleDelete will only be added for
// a key that has no inflight Sets or Merges (Sets have made their way to
// the DB), and no subsequent Sets or Merges will happen until the
// SingleDelete makes its way to the DB.
other.singleDel = other.singleDel || m.singleDel
if other.singleDel {
if other.sets > 1 || other.merges > 0 || other.dels > 0 {
panic(fmt.Sprintf("invalid sets %d or merges %d or dels %d",
other.sets, other.merges, other.dels))
}
}
// Determine if the key is visible or not after the keyMetas are merged.
// TODO(bananabrick): We currently only care about key updates which make it
// to the DB, since we only use key updates to determine if an iterator
// can read a key in the DB. We could extend the timestamp system to add
// support for iterators created on batches.
if other.del || other.singleDel {
other.updateOps = append(
other.updateOps, keyUpdate{true, keyManager.nextMetaTimestamp()},
)
} else {
other.updateOps = append(
other.updateOps, keyUpdate{false, keyManager.nextMetaTimestamp()},
)
}
}
// keyManager tracks the write operations performed on keys in the generation
// phase of the metamorphic test. It makes the assumption that write
// operations do not fail, since that can cause the keyManager state to be not
// in-sync with the actual state of the writers. This assumption is needed to
// correctly decide when it is safe to generate a SingleDelete. This
// assumption is violated in a single place in the metamorphic test: ingestion
// of multiple batches. We sidestep this issue in a narrow way in
// generator.writerIngest by not ingesting multiple batches that contain
// deletes or single deletes, since loss of those specific operations on a key
// are what we cannot tolerate (doing SingleDelete on a key that has not been
// written to because the Set was lost is harmless).
type keyManager struct {
comparer *base.Comparer
// metaTimestamp is used to provide a ordering over certain operations like
// iter creation, updates to keys. Keeping track of the timestamp allows us
// to make determinations such as whether a key will be visible to an
// iterator.
metaTimestamp int
// byObjKey tracks the state for each (writer, key) pair. It refers to the
// same *keyMeta as in the byObj slices. Using a map allows for fast state
// lookups when changing the state based on a writer operation on the key.
byObjKey map[string]*keyMeta
// List of keys per writer, and what has happened to it in that writer.
// Will be transferred when needed.
byObj map[objID][]*keyMeta
// globalKeys represents all the keys that have been generated so far. Not
// all these keys have been written to. globalKeys is sorted.
globalKeys [][]byte
// globalKeysMap contains the same keys as globalKeys. It ensures no
// duplication, and contains the aggregate state of the key across all
// writers, including inflight state that has not made its way to the DB
// yet.The keyMeta.objKey is uninitialized.
globalKeysMap map[string]*keyMeta
// globalKeyPrefixes contains all the key prefixes (as defined by the
// comparer's Split) generated so far. globalKeyPrefixes is sorted.
globalKeyPrefixes [][]byte
// globalKeyPrefixesMap contains the same keys as globalKeyPrefixes. It
// ensures no duplication.
globalKeyPrefixesMap map[string]struct{}
// Using SingleDeletes imposes some constraints on the above state, and
// causes some state transitions that help with generating complex but
// correct sequences involving SingleDeletes.
// - Generating a SingleDelete requires for that key: global.merges==0 &&
// global.sets==1 && global.dels==0 && !global.singleDel && (db.sets==1
// || writer.sets==1), where global represents the entry in
// globalKeysMap[key] and db represents the entry in
// byObjKey[makeObjKey(makeObjID(dbTag, 0), key)], and writer is the
// entry in byObjKey[makeObjKey(writerID, key)].
//
// - We do not track state changes due to range deletes, so one should
// think of these counts as upper bounds. Also we are not preventing
// interactions caused by concurrently in-flight range deletes and
// SingleDelete. This is acceptable since it does not cause
// non-determinism.
//
// - When the SingleDelete is generated, it is recorded as
// writer.singleDel=true and global.singleDel=true. No more write
// operations are permitted on this key until db.singleDel transitions
// to true.
//
// - When db.singleDel transitions to true, we are guaranteed that no
// writer other than the DB has any writes for this key. We set
// db.singleDel and global.singleDel to false and the corresponding sets
// and merges counts in global and db also to 0. This allows this key to
// fully participate again in write operations. This means we can
// generate sequences of the form:
// SET => SINGLEDEL => SET* => MERGE* => DEL
// SET => SINGLEDEL => SET => SINGLEDEL, among others.
//
// - The above logic is insufficient to generate sequences of the form
// SET => DEL => SET => SINGLEDEL
// To do this we need to track Deletes. When db.del transitions to true,
// we check if db.sets==global.sets && db.merges==global.merges &&
// db.dels==global.dels. If true, there are no in-flight
// sets/merges/deletes to this key. We then default initialize the
// global and db entries since one can behave as if this key was never
// written in this system. This enables the above sequence, among
// others.
}
func (k *keyManager) nextMetaTimestamp() int {
ret := k.metaTimestamp
k.metaTimestamp++
return ret
}
// newKeyManager returns a pointer to a new keyManager. Callers should
// interact with this using addNewKey, eligible*Keys, update,
// canTolerateApplyFailure methods only.
func newKeyManager(numInstances int) *keyManager {
m := &keyManager{
comparer: testkeys.Comparer,
byObjKey: make(map[string]*keyMeta),
byObj: make(map[objID][]*keyMeta),
globalKeysMap: make(map[string]*keyMeta),
globalKeyPrefixesMap: make(map[string]struct{}),
}
for i := 1; i <= max(numInstances, 1); i++ {
m.byObj[makeObjID(dbTag, uint32(i))] = []*keyMeta{}
}
return m
}
// addNewKey adds the given key to the key manager for global key tracking.
// Returns false iff this is not a new key.
func (k *keyManager) addNewKey(key []byte) bool {
_, ok := k.globalKeysMap[string(key)]
if ok {
return false
}
keyString := string(key)
insertSorted(k.comparer.Compare, &k.globalKeys, key)
k.globalKeysMap[keyString] = &keyMeta{objKey: objKey{key: key}}
prefixLen := k.comparer.Split(key)
if _, ok := k.globalKeyPrefixesMap[keyString[:prefixLen]]; !ok {
insertSorted(k.comparer.Compare, &k.globalKeyPrefixes, key[:prefixLen])
k.globalKeyPrefixesMap[keyString[:prefixLen]] = struct{}{}
}
return true
}
// getOrInit returns the keyMeta for the (objID, key) pair, if it exists, else
// allocates, initializes and returns a new value.
func (k *keyManager) getOrInit(id objID, key []byte) *keyMeta {
o := makeObjKey(id, key)
m, ok := k.byObjKey[o.String()]
if ok {
return m
}
m = &keyMeta{objKey: makeObjKey(id, key)}
// Initialize the key-to-meta index.
k.byObjKey[o.String()] = m
// Add to the id-to-metas slide.
k.byObj[o.id] = append(k.byObj[o.id], m)
return m
}
// contains returns true if the (objID, key) pair is tracked by the keyManager.
func (k *keyManager) contains(id objID, key []byte) bool {
_, ok := k.byObjKey[makeObjKey(id, key).String()]
return ok
}
// mergeKeysInto merges all metadata for all keys associated with the "from" ID
// with the metadata for keys associated with the "to" ID.
func (k *keyManager) mergeKeysInto(from, to objID) {
msFrom, ok := k.byObj[from]
if !ok {
msFrom = []*keyMeta{}
k.byObj[from] = msFrom
}
msTo, ok := k.byObj[to]
if !ok {
msTo = []*keyMeta{}
k.byObj[to] = msTo
}
// Sort to facilitate a merge.
slices.SortFunc(msFrom, func(a, b *keyMeta) int {
return cmp.Compare(a.String(), b.String())
})
slices.SortFunc(msTo, func(a, b *keyMeta) int {
return cmp.Compare(a.String(), b.String())
})
var msNew []*keyMeta
var iTo int
for _, m := range msFrom {
// Move cursor on mTo forward.
for iTo < len(msTo) && string(msTo[iTo].key) < string(m.key) {
msNew = append(msNew, msTo[iTo])
iTo++
}
var mTo *keyMeta
if iTo < len(msTo) && string(msTo[iTo].key) == string(m.key) {
mTo = msTo[iTo]
iTo++
} else {
mTo = &keyMeta{objKey: makeObjKey(to, m.key)}
k.byObjKey[mTo.String()] = mTo
}
m.mergeInto(k, mTo)
msNew = append(msNew, mTo)
delete(k.byObjKey, m.String()) // Unlink "from".
}
// Add any remaining items from the "to" set.
for iTo < len(msTo) {
msNew = append(msNew, msTo[iTo])
iTo++
}
k.byObj[to] = msNew // Update "to".
delete(k.byObj, from) // Unlink "from".
}
func (k *keyManager) checkForDelOrSingleDelTransition(dbMeta *keyMeta, globalMeta *keyMeta) {
if dbMeta.singleDel {
if !globalMeta.singleDel {
panic("inconsistency with globalMeta")
}
if dbMeta.del || globalMeta.del || dbMeta.dels > 0 || globalMeta.dels > 0 ||
dbMeta.merges > 0 || globalMeta.merges > 0 || dbMeta.sets != 1 || globalMeta.sets != 1 {
panic("inconsistency in metas when SingleDelete applied to DB")
}
dbMeta.clear()
globalMeta.clear()
return
}
if dbMeta.del && globalMeta.sets == dbMeta.sets && globalMeta.merges == dbMeta.merges &&
globalMeta.dels == dbMeta.dels {
if dbMeta.singleDel || globalMeta.singleDel {
panic("Delete should not have happened given SingleDelete")
}
dbMeta.clear()
globalMeta.clear()
}
}
func (k *keyManager) checkForDelOrSingleDelTransitionInDB(dbID objID) {
keys := k.byObj[dbID]
for _, dbMeta := range keys {
globalMeta := k.globalKeysMap[string(dbMeta.key)]
k.checkForDelOrSingleDelTransition(dbMeta, globalMeta)
}
}
// update updates the internal state of the keyManager according to the given
// op.
func (k *keyManager) update(o op) {
switch s := o.(type) {
case *setOp:
meta := k.getOrInit(s.writerID, s.key)
globalMeta := k.globalKeysMap[string(s.key)]
meta.sets++ // Update the set count on this specific (id, key) pair.
meta.del = false
globalMeta.sets++
meta.updateOps = append(meta.updateOps, keyUpdate{false, k.nextMetaTimestamp()})
if meta.singleDel || globalMeta.singleDel {
panic("setting a key that has in-flight SingleDelete")
}
case *mergeOp:
meta := k.getOrInit(s.writerID, s.key)
globalMeta := k.globalKeysMap[string(s.key)]
meta.merges++
meta.del = false
globalMeta.merges++
meta.updateOps = append(meta.updateOps, keyUpdate{false, k.nextMetaTimestamp()})
if meta.singleDel || globalMeta.singleDel {
panic("merging a key that has in-flight SingleDelete")
}
case *deleteOp:
meta := k.getOrInit(s.writerID, s.key)
globalMeta := k.globalKeysMap[string(s.key)]
meta.del = true
globalMeta.del = true
meta.dels++
globalMeta.dels++
meta.updateOps = append(meta.updateOps, keyUpdate{true, k.nextMetaTimestamp()})
if s.writerID.tag() == dbTag {
k.checkForDelOrSingleDelTransition(meta, globalMeta)
}
case *singleDeleteOp:
if !k.globalStateIndicatesEligibleForSingleDelete(s.key) {
panic("key ineligible for SingleDelete")
}
meta := k.getOrInit(s.writerID, s.key)
globalMeta := k.globalKeysMap[string(s.key)]
meta.singleDel = true
globalMeta.singleDel = true
meta.updateOps = append(meta.updateOps, keyUpdate{true, k.nextMetaTimestamp()})
if s.writerID.tag() == dbTag {
k.checkForDelOrSingleDelTransition(meta, globalMeta)
}
case *ingestOp:
// For each batch, merge all keys with the keys in the DB.
for _, batchID := range s.batchIDs {
k.mergeKeysInto(batchID, s.dbID)
}
k.checkForDelOrSingleDelTransitionInDB(s.dbID)
case *applyOp:
// Merge the keys from this writer into the parent writer.
k.mergeKeysInto(s.batchID, s.writerID)
if s.writerID.tag() == dbTag {
k.checkForDelOrSingleDelTransitionInDB(s.writerID)
}
case *batchCommitOp:
// Merge the keys from the batch with the keys from the DB.
k.mergeKeysInto(s.batchID, s.dbID)
k.checkForDelOrSingleDelTransitionInDB(s.dbID)
}
}
func (k *keyManager) eligibleReadKeys() (keys [][]byte) {
return k.globalKeys
}
// eligibleReadKeysInRange returns all eligible read keys within the range
// [start,end). The returned slice is owned by the keyManager and must not be
// retained.
func (k *keyManager) eligibleReadKeysInRange(kr pebble.KeyRange) (keys [][]byte) {
s, _ := slices.BinarySearchFunc(k.globalKeys, kr.Start, k.comparer.Compare)
e, _ := slices.BinarySearchFunc(k.globalKeys, kr.End, k.comparer.Compare)
if s >= e {
return nil
}
return k.globalKeys[s:e]
}
func (k *keyManager) prefixes() (prefixes [][]byte) {
return k.globalKeyPrefixes
}
// prefixExists returns true if a key has been generated with the provided
// prefix before.
func (k *keyManager) prefixExists(prefix []byte) bool {
_, exists := k.globalKeyPrefixesMap[string(prefix)]
return exists
}
func (k *keyManager) eligibleWriteKeys() (keys [][]byte) {
// Creating and sorting this slice of keys is wasteful given that the
// caller will pick one, but makes it simpler for unit testing.
for _, v := range k.globalKeysMap {
if v.singleDel {
continue
}
keys = append(keys, v.key)
}
slices.SortFunc(keys, k.comparer.Compare)
return keys
}
// eligibleSingleDeleteKeys returns a slice of keys that can be safely single
// deleted, given the writer id.
func (k *keyManager) eligibleSingleDeleteKeys(id, dbID objID) (keys [][]byte) {
// Creating and sorting this slice of keys is wasteful given that the
// caller will pick one, but makes it simpler for unit testing.
addForObjID := func(id objID) {
for _, m := range k.byObj[id] {
if m.sets == 1 && k.globalStateIndicatesEligibleForSingleDelete(m.key) {
keys = append(keys, m.key)
}
}
}
addForObjID(id)
if id.tag() != dbTag {
addForObjID(dbID)
}
slices.SortFunc(keys, k.comparer.Compare)
return keys
}
func (k *keyManager) globalStateIndicatesEligibleForSingleDelete(key []byte) bool {
m := k.globalKeysMap[string(key)]
return m.merges == 0 && m.sets == 1 && m.dels == 0 && !m.singleDel
}
// canTolerateApplyFailure is called with a batch ID and returns true iff a
// failure to apply this batch to the DB can be tolerated.
func (k *keyManager) canTolerateApplyFailure(id objID) bool {
if id.tag() != batchTag {
panic("called with an objID that is not a batch")
}
ms, ok := k.byObj[id]
if !ok {
return true
}
for _, m := range ms {
if m.singleDel || m.del {
return false
}
}
return true
}
func opWrittenKeys(untypedOp op) [][]byte {
switch t := untypedOp.(type) {
case *applyOp:
case *batchCommitOp:
case *checkpointOp:
case *closeOp:
case *compactOp:
case *dbRestartOp:
case *deleteOp:
return [][]byte{t.key}
case *deleteRangeOp:
return [][]byte{t.start, t.end}
case *flushOp:
case *getOp:
case *ingestOp:
case *initOp:
case *iterFirstOp:
case *iterLastOp:
case *iterNextOp:
case *iterNextPrefixOp:
case *iterCanSingleDelOp:
case *iterPrevOp:
case *iterSeekGEOp:
case *iterSeekLTOp:
case *iterSeekPrefixGEOp:
case *iterSetBoundsOp:
case *iterSetOptionsOp:
case *mergeOp:
return [][]byte{t.key}
case *newBatchOp:
case *newIndexedBatchOp:
case *newIterOp:
case *newIterUsingCloneOp:
case *newSnapshotOp:
case *rangeKeyDeleteOp:
case *rangeKeySetOp:
case *rangeKeyUnsetOp:
case *setOp:
return [][]byte{t.key}
case *singleDeleteOp:
return [][]byte{t.key}
case *replicateOp:
return [][]byte{t.start, t.end}
}
return nil
}
func loadPrecedingKeys(t TestingT, ops []op, cfg *config, m *keyManager) {
for _, op := range ops {
// Pretend we're generating all the operation's keys as potential new
// key, so that we update the key manager's keys and prefix sets.
for _, k := range opWrittenKeys(op) {
m.addNewKey(k)
// If the key has a suffix, ratchet up the suffix distribution if
// necessary.
if s := m.comparer.Split(k); s < len(k) {
suffix, err := testkeys.ParseSuffix(k[s:])
require.NoError(t, err)
if uint64(suffix) > cfg.writeSuffixDist.Max() {
diff := int(uint64(suffix) - cfg.writeSuffixDist.Max())
cfg.writeSuffixDist.IncMax(diff)
}
}
}
// Update key tracking state.
m.update(op)
}
}
func insertSorted(cmp base.Compare, dst *[][]byte, k []byte) {
s := *dst
i, _ := slices.BinarySearchFunc(s, k, cmp)
*dst = slices.Insert(s, i, k)
}